Table 1           Summary of early technical analysis studies published between 1961 and 1987

 

                       

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

 

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

1. Donchian

    (1960)

 

Copper futures

/ Daily

1959-60

Channel

Not considered

 

$51.50 per round-trip

The current price was compared to the two preceding week’s ranges.  This trading rule generated net gains of $3,488 and $1,390, on margin of $1,000, for a single contract of the December 1959 delivery of copper and the December 1960 delivery, respectively.  

2. Alexander 

    (1961)

S&P Industrials,

Dow Jones Industrials

/ Daily

1897-1959,

1929-59

Filter                           (11 rules from 5.0 to 50%)

Buy & hold

Not adjusted

Trading rules with 5, 6, and 8% filters generated larger gross profits than the B&H (buy-and-hold) strategy.  All the profits were not likely to be eliminated by commissions.  This led Alexander to conclude that there were trends in stock market prices. 

 

3. Houthakker

    (1961)

 

 

Wheat and corn futures

/ Daily

 

1921-39,

1947-56

 

 

Stop-loss order

(11 rules from 0 to 100%)

 

Buy & hold, Sell & hold

 

Not adjusted

 

Most stop-loss orders generated higher profits than the B&H or a sell and hold strategy.  Long transactions indicated better performance than short transactions.

 

4. Cootner (1962)

 

 

45 NYSE stocks

/ Weekly

 

1956-60

 

Moving average

(1/200 days with and without a 5% band)

 

Buy & hold

 

Commissions of 1% per one-way transaction

 

Although net returns from moving average rules were not much different from those from the B&H strategy, long transactions generated higher returns than the B&H strategy.   Moreover, the variance of the trading rule was 30% less than that of the B&H. 

 

5. Gray & Nielsen

    (1963)

 

 

 

Wheat futures

/ Daily

 

1921-43,

1949-62

 

Stop-loss order

(10 rules from 1 to 100%)

 

Buy & hold, Sell & hold

 

Not adjusted

 

When applying stop-loss order rules to dominant contracts, there was little evidence of non-randomness in wheat futures prices.  They argued that Houthakker’s results were biased because he used remote contracts and that post-war seasonality of wheat futures prices was induced by government loan programs. 

 

6. Alexander

    (1964)

 

S&P Industrials

/ Daily

 

1928-61

 

Filter, Formula Dazhi, Formala Dafilt, moving average, and Dow-type formulas

 

Buy & hold

 

Commissions of 2% for each round-trip

 

After commissions, only the largest filter (45.6%) rule beat the B&H strategy by a substantial margin.  Most of the other trading systems earned higher gross profits than filter rules or the B&H strategy.  However, after commissions they could not beat the B&H.

 

7. Smidt (1965a)

 

 

May soybean futures contracts

/ Daily

 

1952-61

 

 

Momentum oscillator (40 rules)

 

Not considered

 

 

$0.36 per bushel per round-trip

 

About 70% of trading rules tested generated positive returns after commissions.  Moreover, half of trading rules returned 7.5% per year or more. 

 

8. Fama & Blume    

    (1966)

 

 

30 individual stocks of the DJIA

/ Daily

 

1956-62

 

 

Filter

(24 rules from 0.5 to 50%)

 

Buy & hold

 

0.1% per round-trip plus other costs

 

After commissions, only 4 of 30 securities had positive average returns per filter.  Even before commissions, filter rules were inferior to the B&H strategy for all but two securities.  Although three small filter rules (0.5, 1.0, and 1.5%) earned higher gross average returns (11.4%-20.9% per year) per security when considering only long positions, net returns after transaction costs were not much different from B&H returns.

 

Table 1 continued.

 

                       

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

 

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

9. Levy (1967a)

 

 

200 NYSE stocks

/ Weekly

 

1960-65

 

Relative strength (Ratios: 1/4 and 1/26 weeks)

 

Geometric average

 

1% per one-way transaction

 

Net returns of several well-performing rules were nearly two or three times the return of the geometric average, although these rules possessed slightly higher standard deviations relative to the geometric average. 

 

10. Levy (1967b)

 

200 NYSE stocks

/ Weekly

 

1960-65

 

Relative strength (Ratio: 1/26 weeks)

 

Not considered

 

1% per one-way transaction

 

Stocks having the historically strongest relative strength showed an average price appreciation of 9.6% over 26 weeks (about 20.1% per year).  An annual price appreciation of all stocks was 12.8%.  In general, stocks that had been both relatively strong and relatively volatile produced higher profits.

 

11. Poole (1967)

 

 

9 exchange rates

/ Daily

 

 

1919-29,

1950-62

 

 

Filter (10 rules from 0.1 to 2%)

 

 

Buy & hold

 

Not adjusted

 

Four of nine exchange rates had average annual gross returns more than 25% for the best filter rules, and three of them (Belgium, France, and Italy) generated returns above 44%.  Filter rules beat the B&H strategy by large differences in returns.

 

12. Van Horne &

      Parker (1967)

 

30 NYSE stocks

/ Daily

 

1960-66

 

Moving average (100, 150, and 200 days with 0, 2, 5, 10, and 15% bands)

 

Buy & hold

 

Commissions charged by members of the NYSE

 

No trading rule earned a total closing balance nearly as large as that generated under the B&H strategy.  Even before transaction costs, gross profits from each moving average rule were less than that from the B&H.

 

 

13. James (1968)

 

232 to 1376 stocks from the CRSP at the Univ. of Chicago

/ Monthly

 

1926-60

 

Moving average

(7 months = 200 days with 2 and 5% bands)

 

Buy & hold

 

Not adjusted

 

Moving average rules could not beat the B&H strategy.  The largest average dollar difference between the moving average rules and the B&H strategy was very small. 

 

 

14. Van Horne &

      Parker (1968)

 

 

 

 

 

30 NYSE stocks

/ Daily

 

1960-66              

 

Non-weighted and exponentially weighted moving averages (200 days with 0, 5, 10, and 15% bands)

 

Buy & hold

 

1% per one-way transaction

 

When applying trading rules to long positions, only 55 of 480 cases (16 different combinations of rules multiplied by 30 stocks) realized profits greater than those from the B&H strategy.  For long plus short positions, a smaller number of trading rules (36 out of 480 cases) outperformed the B&H.   

 

15. Jensen &  

      Benington  

      (1970)

 

29 portfolio samples of 200 NYSE stocks

/ Monthly

 

1931-65

 

Relative strength

(2 rules from Levy (1967a))

 

Buy & hold

 

Actual round lot rate

 

After transaction costs, Levy’s trading rules did not perform better than the B&H strategy.  In fact, after explicit adjustment for the level of risk, the trading rules on average generated net returns less than the risk-adjusted B&H returns. 

 

 

Table 1 continued.

 

                       

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

 

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

16. Stevenson &

      Bear (1970)

 

July corn and soybean futures

/ Daily

 

1957-68

 

Stop-loss order,

filter, and combination of both systems

 

Buy & hold

 

0.5 cents per bushel for both commodities

 

For all systems, a 5% filter rule worked best, which generated larger net profits or greatly reduced losses relative to the B&H strategy.  The filter rule also outperformed B&H for both corn and soybean futures. 

 

17. Dryden

      (1970a)

 

 

U.K. stock indices, Tesco Stores stock

/ Daily

 

 

1962-67,

1962-64

 

Filter (12 rules from 0.1 to 5%)

 

Buy & hold

 

Individual stock: 0.625% per one-way transaction

 

Without transaction costs, filter rules consistently beat the B&H strategy for both indices and an individual stock.  With transaction costs, the returns from the best filter rules were similar to those from the B&H, but long transactions beat the B&H.

 

18. Dryden 

      (1970b)

 

 

15 U.K. stocks

/ Daily

 

1963-64,

1966-67

 

 

Filter (14 rules from 0.2 to 6%)

 

 

Buy & hold

 

Not adjusted

 

There was considerable variation among individual stocks’ returns.  On average, filter returns were less than the corresponding B&H returns except for two smallest filter rules.  However, returns only from long transactions were much higher than the B&H returns.

 

19. Levy (1971)

 

 

 

 

548 NYSE stocks

/ Daily

 

 

1964-69

 

32 forms of a five-point chart pattern

 

Buy & hold

 

2% per round-trip

 

After transaction costs, none of the 32 patterns for any holding period generated profits greater than average purchase or short-sale opportunities.  Even the best-performing pattern produced adjusted relative-to-market returns of -1.1% and -0.1% for one-week and 4-week holding periods, respectively.  

 

20. Leuthold

      (1972)

 

30 live cattle futures contracts

/ Daily

 

1965-70

 

Filter (1, 2, 3, 4, 5, and 10%)

 

Not considered

 

 

Commissions of $36 per round-trip

 

Four of six filters were profitable after transaction costs.  In particular, a 3% filter rule generated an annual net return of 115.8% during the sample period.

 

21. Martell &

      Philippatos

      (1974)

 

 

September wheat and September soybean futures contracts

/ Daily

 

1956-69

(1958-70)*

 

Adaptive filter model and pure information model

 

Buy & hold

/ Optimized trading rules

 

Adjusted but not specified

 

As an optimal filter size for period t, the adaptive model utilizes a filter size which has yielded the highest profits in t-1, subject to some minimum value of the average relative information gain.  The pure information model chooses as an optimal filter size in period t the one with the highest relative average information gain in period t-1.  Both models yielded higher net returns than the B&H only for wheat futures.  However, the variance in net profits was consistently smaller than that of the B&H in both markets.

 

* Years in parentheses indicate out-of-sample periods.

Table 1 continued.

 

                       

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

 

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

22. Praetz (1975)

 

 

Sydney wool futures

/ Daily

 

 

1965-72

 

Filter (24 rules from 0.5 to 25%)

 

Buy & hold

 

Not adjusted

 

For 12 of all 21 contracts of 18-month length and all three 8-year price series, the B&H strategy showed better performance than filter rules, with average differences of 0.1% and 2%, respectively.  For the same data set, in 10 of 24 filters the B&H returns were greater than average filter returns.  Thus, filter rules did not seem to outperform the B&H strategy consistently.

 

23. Martell (1976)

 

 

September wheat and September soybean futures contracts

/ Daily

 

 

1956-69

(1958-70)*

 

Adaptive filter models and pure information model

 

Buy & hold

/ Optimized trading rules

 

Adjusted but not specified

 

A new adaptive model was developed and applied to the same data set as that used in Martell and Philippatos (1974).  The new model selects its optimal filter size for next period based on profitability (e.g., the highest cumulative net profits) and information gain.  Although the model outperformed the previous adaptive model for around 80% of the sample period, it neither indicated any stability with respect to the information constraint nor beat the pure information model that allows a filter size in a particular period to reflect new information.  

 

24. Akemann &

      Keller (1977)

 

Industry groups from S&P 500 Stock Index

/ Weekly

 

1967-75

 

Relative strength

 

S&P 500 Index

 

2% per round-trip

 

The relative strength rule is designed to buy the strongest stock group in a given thirteen-week period and sell it after 52 weeks.  After adjustment for transaction costs, the mean return differential between all 378 possible trials and the market index appeared to be 14.6%, although the differentials were quite volatile. 

 

25. Logue &

      Sweeney (1977)

 

Franc/dollar spot exchange rate

/ Daily

 

1970-74

 

Filter (14 rules from 0.7 to 5%)

 

Buy & hold

 

0.06% per one-way transaction

 

Most trading rules (13 out of 14 rules) outperformed the B&H strategy after considering transaction costs.  Compared to the buy and hold and invest in French government securities strategy, only four filters failed to generate higher profits. 

 

26. Cornell &

      Dietrich

(1978)

 

6 spot foreign currencies (mark, pound, yen, Canadian dollar, Swiss franc, and Dutch guilder)

/ Daily

 

1973-75

 

Filter (13 rules from 0.1 to 5%), and moving average (10, 25, and 50 days with 0.1 to 2% bands)

 

Buy & hold

 

Computed by using the average bid-ask spread for all trades.

 

For the Dutch guilder, German mark, and Swiss franc, the best rules from each trading system generated over 10% annual net returns.  Although the net returns were relatively small (1% to 4%) for the British pound, Canadian dollar, and Japanese yen, they all beat the B&H strategy.  Moreover, since none of the systematic risk (beta) estimates exceeded 0.12, high returns of the three currencies were less likely to be compensation for bearing systematic risk.

 

27. Logue,

      Sweeney, &

      Willett (1978)

 

7 foreign exchange rates

/ Daily

 

1973-76

 

 

Filter (11 rules from 0.5 to 15%)

 

Buy & hold

 

Not adjusted

 

For every exchange rate (the mark, pound, yen, lira, France franc, Swiss franc, and Dutch guilder), profits from the best filter rules exceeded those from the B&H strategy by differences ranging from 9.3% to 32.9%. 

 

* Years in parentheses indicate out-of-sample periods.

Table 1 continued.

 

                       

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

 

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

28. Arnott (1979)

 

 

 

 

 

500 stocks from both the S&P 500 Index and the NYSE Composite Index

/ Weekly

 

1968-77

 

Beta-modified relative strength

 

Not considered

 

Not adjusted

 

Regression results indicated that for the base periods of 1 week to 18 weeks, the correlation between the change in (beta-adjusted) relative strength during the base period and that during any subsequent period was strongly negative.  Hence, careless use of relative strength might lead to serious money loss.

 

29. Dale &

      Workman 

      (1980)

 

90-day T-bill futures at the IMM

/ Daily

 

1976-78

 

Moving average

(11 rules from 5 to 60 days)

 

Not considered

 

$60 per round-trip

 

For each individual contract, the best trading rules generated positive net returns, although the rules did not indicate consistent performances over the sample period.  

 

30. Bohan (1981)

 

 

 

 

 

87 to 110 S&P industry groups

/ Weekly

 

1969-80

 

Relative strength

 

Buy & hold on S&P 500 Index

 

2% per year

 

There was a strong correlation between the performance of the strongest and weakest industry groups in one year and that of the following years, although the performance of the other groups did not have much predictive significance.  For example, quintile 1 portfolio, which consists of the top 20% of industry groups, generated a return of 76% higher than the B&H on the market index, while the market outperformed quintile 5 portfolio by 80%.  

 

31. Solt &

      Swanson

      (1981)

 

 

Gold from London Gold Market and silver from Handy & Harman

/ Weekly

 

1971-79

 

Filter (0.5 to 50%) and moving average (26, 52, and 104 weeks with filters)

 

Buy & hold

 

1.0% per one-way transaction plus 0.5% annual fees

 

For gold, a 10% filter rule outperformed the B&H strategy after adjustment for transaction costs.  However, none of the filter rules dominated the B&H strategy for either gold or silver.  Moving average rules were not able to improve the returns for the filter rules as well. 

 

32. Peterson &

      Leuthold    

      (1982)

 

7 hog futures contracts from CME

/ Daily

 

1973-77

 

Filter (10 rules from 1 to 10% and additional 10 rules from $0.5 to $5)

 

Zero mean profit

 

Not adjusted

 

All 20 filter rules produced considerable mean gross profits.  It seemed that these profit levels exceeded any reasonable commission charges in most cases.  In general, mean gross profits increased with larger filters, as did variance of profits. 

 

33. Dooley &

      Shafer  (1983)

 

9 foreign currencies in the New York market

/ Daily

 

 

1973-81

 

Filter (7 rules from 1 to 25%)

 

Not considered

 

Adjusted but not specified

 

Although results were slightly different for each currency, small filter rules (1, 3, and 5%) generally produced high profits, while larger filter rules showed consistent losses. 

Table 1 continued.

 

                       

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

 

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

34. Brush &

      Boles  (1983)

 

 

 

168 S&P 500 stocks

/ Monthly

 

1967-80,

(two data bases were used for out-of-sample tests)

 

Relative strength

(parameters were optimized on the development data base over 26 separate 6-month test periods)

 

Equal- weighted 168-stock return

/ Optimized models

 

2% per round-trip

 

The top decile annualized excess return of the best model was 7.1% per year over the equal-weighted 168-stock return, after adjustment for risk, dividend yield, and transaction costs.  The model also produced a compounded growth of 15.2% per year after considering dividend yield and transaction costs, compared to 5.9% for the S&P 500.  

 

 

 

35. Irwin & Uhrig

      (1984)

 

 

 

8 commodity futures: corn, cocoa, soybeans, wheat, sugar, copper, live cattle, and live hogs

/ Daily

 

1960-78 (1979-81)*, 1960-68 (1969-72)*, 1973-78 (1979-81)*

 

Channel, moving averages, momentum oscillator

 

 

Zero mean profit

/ Optimized trading rules

 

Doubled commissions to capture bid-ask spread (not specified)

 

Trading rule profits during in-sample periods were substantial and similar across all four trading systems.  Out-of-sample results for optimal trading rules also indicated that during the 1979-81 period most trading systems were profitable in corn, cocoa, sugar, and soybean futures markets.  The trading rule profits appeared to be concentrated in the 1973-81 period.

 

36. Neftci &

      Policano

      (1984)

 

4 futures: copper, gold, soybeans, and T-bills

/ Daily

 

1975-80

 

Moving average

(25, 50, and 100 days) and slope (trendline) method

 

Not considered

 

Not adjusted

 

Trading signals were incorporated as a dummy variable into a regression equation for the minimum mean square error prediction.  Then the significance of the dummy variable was evaluated using F-tests.  Overall, moving average rules indicated some predictive power for T-bills, gold, and soybeans, while the slope method showed mixed results.  

 

37. Tomek &

      Querin  (1984)

 

 

 

3 random price series (each series consists of 300 prices) generated from corn prices for each sample period

/ Daily

 

1975-80,

1973-74,

1980

 

Moving average

(3/10 and 10/40 days)

 

Not considered

 

$50 per round-trip

 

From each of three random prices series, 20 sets of prices were replicated.  The first 20 sets had moderate price variability, the second set large price variability, and the third set drift in prices.  Both trading rules failed to generate positive average net profits for all three groups with an exception of the 10/40 rule for the relatively volatile price group.  The results imply that technical trading rules may earn positive net returns by chance, although they on average could not generate positive net profits.

 

38. Bird (1985)

 

 

Cash and forward contracts of copper, lead, tin, and zinc from London Metal Exchange (LME)

/ Daily

 

1972-82

 

Filter: long positions (and cash profits)

(25 rules from 1 to 25%)

 

Buy & hold

 

1% per round-trip

 

For cash and forward (futures) copper, over 2/3 of filter rules beat the B&H strategy.  Similar results were obtained for lead and zinc but with weaker evidence.  For tin, the results were inconsistent.  Filter rules performed substantially better in the earlier period (1972-77).

* Years in parentheses indicate out-of-sample periods.

Table 1 continued.

 

                       

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

 

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

39. Brush (1986)

 

 

 

 

 

420 S&P 500 stocks

/ Monthly

 

1969-84

 

Relative strength

 

Return of the equal- weighted S&P 500 Index

 

1% per round-trip

 

By avoiding the year-end effect and exploiting beta corrections and the negative predictive power of one-month trends, the best model, which was the generalized least squares beta approach, generated an annual excess return of more than 5% over the equal-weighted S&P 500, after transaction costs. 

 

40. Sweeney

      (1986)

 

 

 

 

Dollar/mark and additional 9 exchange rates

/ Daily

 

 

1973-75 (1975-80)*

 

Filter: long positions

(7 rules from 0.5 to 10%)

 

Buy & hold

/ Optimized trading rules

 

1/8 of 1% of asset value per round-trip

 

Both in- and out-of-sample tests, small filter rules (0.5% to 5%) consistently beat the B&H strategy, and transaction costs did not eliminate the risk-adjusted excess returns of filter rules.  Eight filter rules across 6 exchange rates produced statistically significant excess returns over the B&H in both in- and out-of sample periods.  

 

41. Taylor (1983,

1986)

 

London agricultural futures: cocoa, coffee, and sugar, Chicago IMM currency futures: sterling, mark, and Swiss franc

/ Daily

 

1971-76 (1977-81)*, 1961-73 (1974-81)*, 1974-78 (1979-81)*

 

A statistical price-trend model

 

Buy & hold and interest rate for bank deposit

/ Optimized trading rules

 

1% per round-trip for agricultural futures and 0.2% for currency futures

 

Taylor (1986) adds one more out-of-sample year (i.e., 1981) to the sample period in his 1983’s work.  For sugar, an average net return of the trading rule was higher than that of the B&H strategy by 27% per annum.  For cocoa and coffee, returns from both the trading rule and the B&H were not much different.  Trading gains for currencies during 1979-80 were negligible, but in 1981 all currencies generated substantial gains of around 7% higher than the bank deposit rate.

 

42. Thompson &

      Waller (1987)

 

 

 

 

 

 

 

Coffee and cocoa futures in the NY Coffee, Sugar, and Cocoa Exchange

/ 6 weekly sets of transaction-to-transaction prices for each market

 

 

1981-83

 

Filter

(for coffee, 5¢ through 35¢ in multiples of 5¢ per 100 lb; for cocoa, $1 through $7 per metric ton)

 

Not considered

 

Estimated execution costs

 

For both nearby and distant coffee and cocoa contracts, filter rules generated average profits per trade per contract substantially lower than estimated execution costs per contract in all cases in which profits were statistically significantly greater than zero.  The estimated execution costs per trade per contract were $32.25 (nearby) and $69.75 (distant) for coffee futures contracts and $12.60 (nearby) and $21.80 (distant) for cocoa futures contracts. 

* Years in parentheses indicate out-of-sample periods.

 

 

Table 2          Categories for modern technical analysis studies

 

 

Category

 

 

Number of studies

 

Representative study

 

Transaction costs

 

 

Risk adjustment

        Criteria

 

Trading rule

optimization

­

 

Out-of-sample tests

 

 

Statistical tests

 

 

Data snooping addressed

           

Distinctive features

 

Standard

 

23

 

Lukac, Brorsen, & Irwin (1988)

 

 

 

 

 

 

 

Conduct parameter optimization and out-of-sample tests.

 

Model-based

bootstrap

 

21

 

Brock, Lakonishok, & LeBaron (1992)

 

 

 

 

 

 

 

 

Use model-based bootstrap methods for statistical tests.  No parameter optimization and out-of-sample tests conducted.

 

Genetic programming

 

11

 

Allen & Karjalainen (1999)

 

 

 

 

 

 

 

Use genetic programming techniques to optimize trading rules.

 

Reality Check

 

3

 

Sullivan, Timmermann, & White (1999)

 

 

 

 

 

 

 

Use White’s Reality Check Bootstrap methodology for optimization and statistical tests.

 

Chart patterns

 

11

 

Chang & Osler (1999)

 

 

 

 

 

 

 

Use recognition algorithms for chart patterns.

 

Nonlinear

 

 

7

 

Gençay (1998a)

 

 

 

 

 

 

 

Use nearest neighbors and/or feedforward network regressions to generate trading signals.

 

Others

 

16

 

Neely (1997)

 

 

 

 

 

 

 

Most studies in this category lack trading rule optimization and out-of-sample tests, and do not address data-snooping problems.

 

 

Table 3           Summary of standard technical analysis studies published between 1988 and 2004

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

1. Lukac,

    Brorsen,

    & Irwin (1988)

 

 

12 futures from various exchanges: agriculturals, metals, currencies, and interest rates

/ Daily

 

1975-83  (1978-84)

 

 

12 systems

(3 channels,

3 moving averages, 3 oscillators,

2 trailing stops, and a combination)

 

Zero mean profit

/ Optimized trading rules

 

$50 and $100 per round-trip

 

Out-of-sample results indicated that 4 of 12 systems generated significant aggregate portfolio net returns and 8 of the 12 commodities earned statistically significant net returns from more than one trading system.  Mark, sugar, and corn markets appeared to be most profitable during the sample period.  In addition, Jensen test confirmed that the same four trading systems having large net returns still produced significant net returns above risk. 

 

2. Lukac &

    Brorsen (1989)

 

 

15 futures from various exchanges: agricultural commodities, metals, currencies, and interest rates

/ Daily

 

1965-85

(various)

 

Channel and

directional movement (both systems had 12 parameters ranging 5 days to 60 days in increments of 5)

 

Buy & hold

/ Optimized trading rules

 

$100 per round-trip

 

Technical trading rule profits were measured based on various optimization methods, which included 10 re-optimization strategies, one random strategy, and 12 fixed parameter strategies.  The two trading systems generated portfolio mean net returns significantly greater than the B&H strategy.  However, the trading systems yielded similar profits across different optimization strategies and even different parameters.  Thus, the parameter optimization appeared to have little value.      

 

3. Sweeney &

    Surajaras

    (1989)

 

 

 

An equally-weighted portfolio and a variably-weighted portfolio of currencies

/ Daily

 

Prior 250- to 1400-day prices

(1980-86)

 

Filter, single moving average, double moving average, and the best system

 

Buy & hold

/ Optimized trading rules

 

Adjusted but not specified

 

Most trading systems generated risk-adjusted mean net profits after transaction costs, and the single moving average rule performed best.  The variably-weighted portfolio approach generally outperformed the equally-weighted approach.  Changing neither parameters for each trading system on a yearly basis nor amounts of data used to select optimal parameters seem to improve trading profits. 

 

4. Taylor & Tari

    (1989)

 

IMM currency futures: pound, mark, and Swiss franc; London agricultural futures: cocoa, coffee, and sugar

/ Daily

 

1974-78

(1979-87);

(1982-85)

 

 

A statistical price-trend model

 

Buy & hold,

Zero mean profit

/ Optimized trading rules

 

Currency futures: 0.2% per round-trip; Agricultural futures: 1%

 

 

During the out-of-sample period, 1979-87, the trading rule earned aggregate mean net return of 4.3% per year for three currency futures.  The mark was the most profitable contract (5.4% per year).  From 1982-85, the trading rule generated a mean net return of 4.8% for cocoa, -4.26% for coffee, and 18.8% for sugar, outperforming the B&H strategy for cocoa and sugar futures.    

 

5. Lukac &

    Brorsen (1990)

 

30 futures from various exchanges: agriculturals, metals, oils, currencies, interest rates, and S&P 500

/ Daily

 

1975-85

(1976-86)

 

23 systems (channels, moving averages, oscillators, trailing stops, point and figure, a counter-trend, volatility, and combinations)

 

 

Zero mean profit

/ Optimized trading rules

 

$50 and $100 per round-trip

 

Only 3 of 23 trading systems had negative mean monthly portfolio net returns after transaction costs, and 7 of 23 systems generated net returns significantly above zero at 10% level.  Most of the trading profits appeared to be made over the 1979-80 period.  In the individual commodity markets, currency futures produced the highest returns, while livestock futures yielded the lowest returns. 

 

Table 3 continued.

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

6. Taylor (1992)

 

 

 

4 currency futures from IMM of the CME:  pound, mark, yen, and Swiss franc

/ Daily

 

1977-87

(1982-87)

 

 

3 technical trading systems (filter, channel, moving average), 2 statistical price-trend models

 

Buy & hold

/ Optimized trading rules

 

0.2% per round-trip

 

All trading rules outperformed the B&H strategy across all currency futures.  Among trading rules, three technical trading systems and a revised statistical trend model generated statistically significant and much higher mean net returns (3.0% to 4.0%) than that (2.0%) of the original price-trend model for most currencies.  These returns could not be explained by nonsynchronous trading or time-varying risk premia. 

 

7. Farrell &

    Olszewski

    (1993)

 

 

 

 

S&P 500 futures

/ Daily

 

1982-90

(1989-90)

 

A nonlinear trading strategy based on ARMA (1,1) model and 3 trend-following systems (channel and volatility systems)

 

Buy & hold

/ Optimized trading rules

 

0.025% per round-trip

 

Although the nonlinear trading strategy were slightly more profitable than the B&H strategy, the result was statistically insignificant.  For the in-sample period, the nonlinear optimal trading strategy was more profitable than the B&H by nearly 5%, while for the out-of-sample period, the trading strategy was better by 3%.  Meanwhile, the three trend following strategies were more profitable than the nonlinear trading strategy by around 5% to 11% during the out-of-sample period, depending on the trading strategy. 

 

8. Silber (1994)

 

 

12 futures markets: foreign currencies, short-term interest rates, metals, oil, and S&P 500

/ Daily

 

 

1979

(1980-91)

 

 

Moving average

(short averages: 1 day to 15 days; long averages: 16 to 200 days)

 

 

Buy & hold (& roll over)

/ Optimized trading rules

 

Bid-ask spreads per round-trip (2 ticks for crude oil and gold; 1 tick for the rest of contracts)

 

After transaction costs, average annual net returns were positive for all contracts but gold, silver, and the S&P 500.  In particular, most currency futures earned higher net profits (1.9% to 9.8%).  For those profitable markets, moving average rules beat the B&H strategy except for 3-month Eurodollars.  Test results using a Sharpe ratio criterion were similar.  Hence, trading profits appeared to be robust to transaction costs and risk.  Central bank intervention is one of possible explanations for the trading profits.

 

9. Taylor (1994)

 

 

 

 

4 currency futures from IMM: pound, mark, yen, and Swiss franc

/ Daily

 

1980-all previous contracts (1982-90)

 

Channel

 

Zero mean profits

/ Optimized trading rules

 

0.2% per one-way transaction

 

For price series generated by ARIMA(1,1,1) model, channel rules correctly identified the sign of conditional expected returns with around 60% probability.  During 1982-90, optimal channel rules produced an average net return of 6.9% per year.  The t-test indicated that the return was significant at the 2.5% level.  The best trading opportunities occurred for 1985-87. 

 

10. Menkhoff &

      Schlumberger

      (1995)

 

3 spot exchange rates: mark/dollar, mark/yen, and mark/pound

/ Daily

 

1981-91,

1981-85

(1986-91)

 

 

 

 

Oscillator (33 moving averages)

and momentum (10 rules from 5 to 40 days)

 

 

 

Buy & hold

/ Optimized trading rules

 

0.0008 DM for 1$; 0.0017 DM for 1 yen; 0.003 DM for 1 BP per round-trip

 

 

During the out-of-sample period, 84% out of 129 technical trading rules tested outperformed the B&H strategy across exchange rates, after adjustment for transaction costs and risk.  However, superiority of optimal trading rules during the in-sample period deteriorated in the out-of-sample period, even though they still outperformed the B&H strategy. 

 

 

Table 3 continued.

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

11. Lee &

      Mathur

      (1996a)

 

6 European currency spot cross-rates

/ Daily

 

1988-92

(1989-93)

 

Moving average

(short moving averages: 1 day to 9 days; long moving averages: 10, 15, 20, 25, and 30 days)

 

Zero mean profits

/ Optimized trading rules

 

0.1% per round-trip

 

Results of in-sample tests indicated that the trading rules did not yield significantly positive returns for all cross rates but yen/mark and yen/Swiss franc (11.5% and 8.8% per year, respectively).  Out-of-sample results were even worse.  Most cross rates earned negative trading returns, although long positions for the yen/mark produced marginally significant positive returns.

 

12. Lee &

      Mathur

      (1996b)

 

 

10 spot cross-rates

/ Daily

 

1988-92

(1989-93)

 

 

Moving average

(short moving averages: 1 day to 9 days; long moving averages: 10, 15, 20, 25, and 30 days) and channel (2 to 50 days)

 

Zero mean profits

/ Optimized trading rules

 

0.1% per round-trip

 

During in-sample periods, moving average rules in general produced negative or statistically insignificantly positive net returns except the mark/yen (11.5% per year) and the Swiss franc/yen (8.8% per year).  Similar results were found for channel rules.  During out-of-sample periods, overall returns of the trading rules were negative or statistically insignificantly positive.  Only for the mark/lira, both long positions of moving average rules and channel rules generated statistically significant profits.

 

13. Szakmary &

      Mathur

      (1997)

 

 

5 IMM foreign currency futures and spots: mark, yen, pound, Swiss franc, and Canadian dollar

/ Daily

 

1977-90

(1978-91)

 

Moving average

(short moving averages: 1 day to 9 days; long moving averages: 10, 15, 20, 25, and 30 days)

 

Zero mean profits

/ Optimized trading rules

 

0.1% per round-trip

 

In-sample results indicated that moving average rules generated both statistically and economically significant returns for all currency futures but the Canadian dollar.  Similar results were reported for both out-of-sample data (annual net returns ranged from 5.5% to 9.6%) and spot rates.  Further analyses showed that the moving average rule profits resulted from the central bank’s “leaning against the wind intervention.”

 

14. Goodacre,

      Bosher, &  

      Dove (1999)

 

 

 

254 companies in the FTSE 350 Index and 64 option trades in the U.K.

/ Daily

 

Prior 200 days

(1988-96)

 

CRISMA (combination system of Cumulative volume, RelatIve Strength, and Moving Average)

 

FTSE All Share Index

/ Optimized parameters

 

0 to 2% per round-trip

 

The CRISMA trading system generated annualized profits ranging 6.9% to 19.3% depending on transaction costs, while an annualized return on the FTSE All Share Index over the same time period was 14.0%.  When adjusted for market movements and risk, however, mean excess returns for nonzero levels of transaction costs were significantly negative.  Moreover, performance of the trading system was not stable over time.  With option trading, the system generated mean return of 10.2% per trade even in the presence of maximum retail costs, but only 55% of trades were profitable. 

 

15. Kwan, Lam,

      So, & Yu

      (2000)

 

 

Hang Seng Index Futures

/ Daily

 

1986-97

(1990-98)

 

A statistical price-trend model

 

Buy & hold /

Optimized parameters

 

0.4 to 0.5% per one-way transaction

 

The price-trend model performed poorer than the B&H strategy in the periods 1991-93 and 1995-96 when the market was bullish.  However, the trading rule produced larger profits than the B&H in the years, 90, 94, 97, and 98 when the market became up and down.  Across all years and transaction costs considered, an average net return (10.1%) of the trading rule was slightly smaller than that (13.5%) of the B&H strategy.

 

Table 3 continued.

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

16. Maillet &

      Michel

      (2000)

 

 

12 exchange rates (combinations of U.S. dollar, mark, yen, pound, and France franc)

/ Daily

 

1974-79

(1979-96)

 

Moving average

(short moving averages: 1 day to 14 days; long moving averages: 15 to 200 days)

 

Zero mean profits, buy & hold

/ Optimized trading rules

 

Not adjusted

 

Optimized moving average rules generated statistically significant returns and outperformed the corresponding B&H strategies with the exception of the mark/franc rate.  Bootstrap tests generally confirmed the results with the rejection of higher returns only in 4 out of 12 rates: the mark/dollar, mark/franc, yen/dollar, and yen/franc.  Moreover, riskiness of both moving average rules and the B&H strategy, which was measured by their standard deviations, appeared to be not much different. 

 

17. Taylor (2000)

 

 

 

 

 

1) Financial Times (FT) All-Share index; 2) UK 12-share index; 3) 12 UK stocks; 4) FT 100 index and index futures; 5) DJIA index; 6) S&P 500 index and index futures

/ Daily

 

1), 2), and 3): 1972-91;

4): 1985-94;

5): 1897-1988;

6): 1982-92

 

 

Moving average 

(short moving averages: 1, 2, and 5 days; long moving averages: 50, 100, 150, and 200, with and without a 1% band)

 

 

/ Parameters are optimized for the DJIA data from 1897 to 1968.

 

Not adjusted

 

 

The results of optimized moving average rules indicated that differences of mean returns between buy and sell positions were substantially positive and statistically significant for the FTA index, all versions of the 12-share index, 4 of the 12 UK firms, and the DJIA index for 3 out of 5 subperiods.  No significant results were found for the FTSE 100 and S&P 500 indices.  Buy positions also appeared to have lower standard deviations than sell positions for all but two series.  An average breakeven one-way transaction cost across all data series was 0.35%.  In particular, for the DJIA index, a trading rule (a 5/200 moving average rule) optimized over the 1897-1968 period produced a breakeven one-way transaction cost of 1.07% during the 1968-88 period.   

 

18. Goodacre &

      Kohn-

      Spreyer

      (2001)

 

 

 

 

A random sample of 322 companies from the S&P 500

/ Daily

 

Prior 200 days

(1988-96)

 

CRISMA (combination system of Cumulative volume, RelatIve Strength, and Moving Average)

 

The S&P 500 Index

/ Optimized parameters

 

0 to 2% per round-trip

 

The CRISMA system generated annualized profits ranging 6.2% to 17.6% depending on transaction costs, while the annualized return on the S&P 500 Index over the same time period was 14.2%.  However, when adjusted for market movements and risk, mean excess returns for nonzero levels of transaction costs were significantly negative across all return-generating models.  Moreover, the results were not stable over time, although trades on larger firms generally performed better than small ones.

 

19. Lee,

      Gleason,

      & Mathur

      (2001)

 

 

 

 

13 Latin American spot currencies

/ Daily

 

1992-99 (various periods from data available)

 

Moving average

(short moving averages: 1 day to 9 days; long moving averages: 10 to 30 days) and channel (2 to 50 days)

 

 

Zero mean profits

/ Optimized trading rules

 

0.1% per round-trip

 

Out-of-sample results showed that moving average rules generated significantly positive returns for currencies of four countries: Brazil, Mexico, Peru, and Venezuela.  Channel rules also produced significant profits for the same currencies except that of Peru.  When only long positions were considered, there was a marginal improvement to five and four currencies for moving average rules and channel rules, respectively. 

Table 3 continued.

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

20. Lee, Pan, &

      Liu (2001)

 

9 exchange rates from Asian countries

 

1988-94 (1989-95)

 

The same trading rules as in Lee, Gleason, & Mathur (2001)

 

Zero mean profits

/ Optimized trading rules

 

0.1% per round-trip

 

Out-of-sample tests indicated that four exchange rates from Korea, New Zealand, Singapore, and Taiwan yielded positive profits for both moving average rules and channel rules.  However, these profits were not significantly different from zero, except that of the Taiwan dollar. 

 

21. Martin

      (2001)

 

 

 

 

12 currencies in developing countries

/ Daily

 

1/92-6/92

(7/92-6/95)

 

 

Moving average

(short moving averages: 1 day to 9 days; long moving averages: 10 to 30 days)

 

Short-selling strategy

/ Optimized trading rules

 

0.5% per one-way transaction

 

Out-of-sample, moving average rules generated positive mean net returns in 10 of 12 currencies, and the returns were greater than 0.14% daily (35% per year) in 5 currencies.  However, Sharpe ratios indicated that moving average rules did not generate superior returns on a risk-adjusted basis. 

 

22. Skouras

      (2001)

 

 

 

 

 

Dow Jones Industrial Average (DJIA)

/ Daily

 

1962-86

(1962-86)

 

Moving average

(2 to 200 days with bands of 0, 0.5, 1, 1.5, and 2%)

 

Buy & hold

/ Optimized trading rules

 

Various levels from 0 to 0.1% per one-way transaction

 

Out-of-sample returns were estimated on a daily basis.  Time-varying estimated rules (by an Artificial Technical Analyst) outperformed various fixed moving average rules employed by Brock et al. (1992) as well as the B&H strategy.  When considering transaction costs, however, mean returns from the optimized trading rule were higher than the B&H mean return only after transaction costs of less than 0.06%. 

 

23. Olson (2004)

 

18 exchange rates

/ Daily

 

5-year in-sample period from 1971-2000 (1976-2000)

 

 

Moving average

(short moving averages: 1 day to 12 days; long moving averages: 5 to 200 days) 

 

Buy & hold

/ Optimized trading rules

 

 

0.1% per round-trip

 

Out-of-sample results indicated that risk-adjusted trading profits for individual currencies and an equal-weighted 18-currency portfolio declined over time.  For the 18-currency portfolio, annualized risk-adjusted returns decreased from an average of over 3% in the late 1970s and early 1980s to about zero percent in the late 1990s.  Overall, profits of moving average rules in foreign exchange markets have declined over time. 

 

 

 

Table 4           Summary of model-based bootstrap technical analysis studies published between 1988 and 2004

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

1. Brock,

    Lakonishok,  &

    LeBaron (1992)

 

 

 

 

Dow Jones Industrial Average (DJIA)

/ Daily

 

1897-1986

 

Moving averages

(1/50, 1/150, 5/150, 1/200, and 2/200 days with 0 and 1% bands) and trading range breakout (50, 150 and 200 days with 0 and 1% bands)

 

Unconditional 1- and 10-day returns

 

Not adjusted

 

Before transaction costs, buy (sell) positions across all trading rules consistently generated higher (lower) mean returns than unconditional mean returns, and these results were highly significant in most cases.  For example, a mean buy return from variable moving average rules was about 12% per year and a mean sell return was about -7%.  Moreover, the buy returns were even less volatile than the sell returns.  Simulated series from a random walk with a drift, AR (1), GARCH-M, and EGARCH models using a bootstrap method could not explain returns and volatility of the actual Dow series.  

 

2. Levich &

    Thomas

    (1993)

 

 

 

5 IMM currency futures: mark, yen, pound, Canadian dollar, and Swiss franc

/ Daily

 

1976-90

 

Filters (0.5, 1, 2, 3, 4, and 5%) and moving average (1/5, 5/20, 1/200 days)

 

Buy & hold

 

0.025% and 0.04% per one-way transaction

 

After adjustment for transaction costs and risk, every filter rule and moving average rule generated substantial positive mean net returns for all currencies but the Canadian dollar.  Moreover, the results of the bootstrap simulation indicated that, for both trading systems, the null hypothesis that there is no information in the original time series was rejected in 25 of 30 cases. 

 

3. Bessembinder

    & Chan (1995)

 

 

Asian stock indices: Hong Kong, Japan, Korea, Malaysia, Thailand, and Taiwan

/ Daily

 

1975-91

 

The same trading rules as in Brock et al. (1992)

 

Buy & hold

 

0.5, 1, and 2% per round-trip

 

Across all markets and trading rules tested, average mean returns on buy days exceeded those on sell days by 26.8% per year, and an average break-even round-trip transaction cost for the full sample was 1.57%.  In particular, technical signals generated by the U.S.  markets appeared to have substantial forecast power for returns in the Asian markets.  Overall, trading rules generated higher net profits (12.2% to 21.2% per year) in the Malaysia, Thailand, and Taiwan stock markets.

 

4. Hudson,

    Dempsey,

    & Keasey

    (1996)

 

 

Financial Times Industrial Ordinary Index (FT30) in the U.K.

/ Daily

 

1935-94

 

The same trading rules as in Brock et al. (1992)

 

Unconditional mean returns

 

More than 1% per round-trip for large investing institutions

 

Before transaction costs, buy (sell) positions across all trading systems consistently generated higher (lower) returns than unconditional returns.  However, an extra return per round-trip transaction averaged across all systems appeared to be about 0.8%, which was relatively smaller than the round-trip transaction costs of 1%. 

 

5. Kho (1996)

 

 

 

 

4 currency futures from IMM: pound, mark, yen, and Swiss franc

/ Weekly

 

 

1980-91

 

Moving average

(1/20, 1/30, 1/50, 2/20, 2/30, 2/50 weeks with bands of 0 and 1%)

 

Unconditional weekly mean return, Univariate GARCH-M

 

Not adjusted

 

Initially, moving average rules generated substantial mean returns between 9.9% and 11.1% per year from buy signals.  These trading returns could not be explained by the empirical distribution of the univariate GARCH-M model as well as transaction costs or serial correlations in futures returns.  However, the returns appeared to be insignificant when time-varying risk premia, which were estimated from a general model of the conditional CAPM, were taken into account.    

 

 

Table 4 continued.

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

6. Raj & Thurston

    (1996)

 

 

Hang Seng Futures Index of Hong Kong

/ Daily

 

 

1989-93

 

The same trading rules as in Brock et al. (1992), without 1/150 and 2/200 moving average rules

 

Unconditional mean returns

 

Not adjusted

 

Without considering transaction costs, average buy returns generated from both trading systems were much higher than the unconditional one-day mean.  In particular, the trading range breakout system generated significantly higher annual returns (457% to 781%) in four out of six rules relative to that (39%) of the B&H strategy.  On the other hand, average sell returns obtained from both systems were negative.

 

7. Mills (1997)

 

 

Financial Times–Institute of Actuaries 30 (FT30) index in the London Stock Exchange

/ Daily

 

 

1935-94:

1935-54,

1955-74,

1975-94

 

The same trading rules as in Brock et al. (1992)

 

Unconditional mean daily return

 

Not adjusted

 

For moving average rules, each mean daily buy-sell return difference (0.081% and 0.097%) for 1935-54 and 1955-74 was much greater than corresponding unconditional mean returns (0.013% and 0%).  For the latest subperiod, 1975-94, however, the mean buy-sell difference was insignificantly different from the unconditional return.  Trading range breakout rules showed similar results.  None of simulated series generated by AR-ARCH bootstraps earned mean buy-sell differences larger than the actual difference. 

 

8. Bessembinder

    & Chan (1998)

 

 

Dow Jones Industrial Average (DJIA)

/ Daily

 

1926-91:

1926-43,

1944-59,

1960-75,

1976-91

 

The same trading rules as in Brock et al. (1992)

 

Buy & hold

 

Various estimates for NYSE stocks

 

 

 

The DJIA data in this study includes dividend payments.  Over the full sample period, an average buy-sell return difference across all 26 trading rules was 4.7%, generating a break-even one-way transaction cost of 0.39%.  However, break-even transaction costs have declined over time with 0.22% for the most recent subperiod (1976-91).  It was compared with an estimated transaction cost of 0.25%. 

 

9. Ito (1999)

 

 

 

6 national equity market indices (Japan, U.S. Canada, Indonesia, Mexico, Taiwan), Dow Jones index, Nikkei index futures

/ Daily

 

1980-96 for developed markets,

1988-96 for emerging markets

 

The same trading rules as in Brock et al. (1992)

 

Buy & hold

 

Nikkei index futures: 0.11% per round-trip; other equity indices: 0.69-2.21%

 

After transaction costs, technical trading rules outperformed the B&H strategy for all indices but U.S.  indices, and generated higher profits for emerging markets (Indonesia, Mexico, Taiwan) than for developed markets.  The trading profits could not be explained by nonsynchronous trading.  However, some conditional asset pricing models (in particular, the asset pricing model under mild segmentation) were able to explain trading rule profits for Japan, the U.S., the second subperiod of Canada, and Taiwan stock indices.  These results suggest that technical trading profits were a fair compensation for risk of trading rules.  

 

10. LeBaron

      (1999)

 

 

2 foreign currencies from the London close: mark and yen

/ Daily and weekly

 

1979-92

 

Moving average (1/150 days or 1/30 weeks)

 

Sharpe ratio for buying and holding on U.S. stock portfolios

 

Commissions (0 to 0.5%) and bid-ask spread (0.15%) per round-trip

 

Mean returns of the trading rule for the two currencies were statistically significantly different from zero.  Their Sharpe ratios (0.60 to 0.98) were also higher than those (0.3 or 0.4) for the B&H on U.S.  stock portfolios even after adjustment for a transaction cost of 0.1% per round-trip.  In general, interest differentials and transaction costs did not alter the result greatly.  However, trading returns were dramatically reduced when active intervention periods of the Federal Reserve were eliminated. 

 

Table 4 continued.

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

11. Ratner & Leal

      (1999)

 

 

10 equity indices in Asia and Latin America

/ Daily

 

 

1982-95

 

Moving average

(1/50, 1/150, 5/150, 1/200, and 2/200 days with bands of zero and one standard deviation)

 

Buy & hold

 

Various costs from 0.15 to 2.0% per one-way transaction

 

After transaction costs, 21 out of 100 trading rules that were applied to the 10 indexes generated statistically significant returns (18.2% to 32.1% per year), with the profitability concentrated in four markets: Mexico, Taiwan, Thailand, and the Philippines.  When statistical significance was ignored, however, 82 out of the 100 rules appeared to have forecasting ability in emerging markets. 

 

12. Coutts &

      Cheung  

      (2000)

 

Hang Seng Index on the Hong Kong Stock Exchange

/ Daily

 

1985-97

 

 

The same trading rules as in Brock et al. (1992)

 

Unconditional mean returns

 

Not adjusted

 

Across all trading rules tested, buy (sell) signals generated significantly higher (lower) mean returns than unconditional mean returns.  In particular, buy (sell) signals of the trading range breakout system earned substantial average 10-day cumulative return of 1.6% (-5%), which was higher (lower) than that of the moving average system.

 

13. Parisi &

      Vasquez

      (2000)

 

 

 

Santiago stock index

/ Daily

 

1987-98

 

The same trading rules as in Brock et al. (1992)

 

Unconditional mean returns

 

1% per one-way transaction

 

Across trading rules, mean returns on buy signals were consistently higher than those on sell signals or unconditional mean returns.  In fact, sell signals yielded negative mean returns for most trading rules.  Although variable-length moving average rules generated significant returns, it was unlikely that these rules were profitable if high transaction costs were taken into account.  

 

14. Raj (2000)

 

 

Yen and mark traded in Singapore International Monetary Exchange

/ Intra-daily

 

 

01/1992-12/1993

 

Filter, moving average, and channel

 

Buy & hold

 

0.04% per one-way transaction

 

None of technical trading rules except one rule (2/200 moving average rule with a 1% band) generated statistically significant returns after adjustment for transaction costs and risk.  However, some trading rules appeared to produce economically significant returns.  For instance, for the mark a 1/50 moving average rule with a 1% band generated a risk-adjusted net return of 8.8% over the two-year period. 

 

15. Gunasekarage

      & Power

      (2001)

 

4 South Asian stock indices: Bombay, Colombo, Dhaka, and Karachi stock exchanges

/ Daily

 

1990-2000

 

Moving averages

(1/50, 1/100, 1/150, 1/200, 2/100, 2/150, 2/200, 5/200, and 1/50 with 1% band)

 

Buy & hold

 

Not adjusted

 

For variable moving average rules, buy signals generated positive returns of more than 44.2% per year and sell signals generated negative returns of less than -20.8% per year.  These returns, on average, were significantly different from the B&H returns.  Similar results were obtained for fixed-length moving average rules with 10-day holding periods. 

 

Table 4 continued.

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

16. Day & Wang

      (2002)

 

Dow Jones Industrial Average (DJIA)

/ Daily

 

1962-96

 

Moving average

(1/50 and 1/150 days with 0 and 1% bands) and

trading range breakout (50 and 150 days with 0 and 1% bands)

 

Buy & hold

 

0.05% per one-way transaction

 

Variable-length moving average rules generated daily excess returns of more than 0.027% over the B&H strategy for 1962-86, and all the returns were statistically significant.  For closing levels of the DJIA that were estimated to reduce the effects of nonsynchronous trading, the trading rules also outperformed the B&H, although returns were reduced relative to previous ones and not all were statistically significant.  For 1987-96, however, the performance of the trading rules was inferior to the B&H strategy in most cases. 

 

17. Kwon & Kish

      (2002)

 

 

 

 

 

The NYSE value-weighted index

/ Daily

 

1962-96: 1962-72,

1973-84,

1985-96

 

 

Moving average,

combination of   moving average and momentum,

and combination of moving averages for price and volume

 

Unconditional mean returns

 

Not adjusted

 

Combination moving average rules of price and volume generated the highest daily average return of 0.13% over the full sample period.  Across all subperiods but the recent 1985-96 period, returns of the trading system were statistically significantly different from unconditional mean returns.  Similar results were obtained for the other two trading systems.  Simulated series from three popular models (random walk, GARCH-M, and GARCH-M with instrument variable) could not explain returns and volatility of the technical trading systems.  

 

18. Neely (2002)

 

4 foreign exchange rates: mark, yen, Swiss franc, and Australia dollar

/ Intra-daily and daily

 

1983-98

 

Moving average

(1/150)

 

Not considered

 

Not adjusted

 

With daily data, the moving average rule generated positive annual mean returns for all series ranging from 2.4% for the Australian dollar to 8.7% for the yen.  However, when intervention periods of central banks were removed, the trading rule returns were greatly reduced, ranging from –2.3% to 4.5%.  With intra-daily data, the highest US, Swiss, and German excess returns appeared to precede business hours and thus precede intervention.  Hence, intervention was less likely to be a cause that generated trading rule profits.

 

19. Saacke (2002)

 

 

Dollar/mark exchange rate in the New York market

/ Daily

 

1979-94

 

Moving average

(2 to 500 days)

 

Not considered

 

0.05% per round-trip

 

Moving average rules below 170 days earned positive net returns.  Bootstrapping simulations based on a random walk with drift and a GARCH model could not account for the size of trading rule returns.  Moving average rules appeared to be highly profitable on days when central banks intervened.  However, since trading rule returns in periods that neither coincided with nor were preceded by interventions were also sizable, interventions did not seem to be the only cause of the trading rule profitability. 

 

Table 4 continued.

 

 

              Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

20. Fang & Xu

      (2003)

 

 

 

3 Dow Jones Indexes

(Industrial, Transportation, and Utilities Averages)

/ Daily

 

1896-1996

 

Moving average,

time series models, and

combination of moving average and time series models

 

Buy & hold

 

Various estimates

 

When the market was bullish (bearish), technical trading rules performed in general better (worse) than trading strategies based on time series models.  When a monthly interest rate of 0.30% was assumed over the full sample period, combination rules produced average break-even transaction costs of about 1.01%, 1.96%, and 1.76% for the Industrial, Transportation, and Utilities Averages, respectively, with non-synchronous trading adjustment.  These figures appeared to be substantial improvement on those of moving average rules (0.60%, 0.84%, and 0.80%, respectively). 

 

21. Sapp (2004)

 

 

 

 

 

 

 

Mark and yen

/ Daily

 

1975-1998

 

Moving average

 

Sharpe ratio for S&P500

 

Bid-ask spread

 

During the 1980-94 period, moving average rules generated statistically and economically significant returns.  Positive but insignificant returns after 1995 seemed to be related with a decrease in central bank intervention activities.  Transaction costs did not affect technical trading returns except for a few short-term trading rules.  Over the 1980-98 period, annualized Sharpe ratios for a 150-day trading rule and investing in the S&P500 were 0.65 and 0.49, respectively.  However, a preliminary analysis using an international CAPM indicated that the hypothesis that there was a time-varying risk premium in the technical trading returns correlated with central bank interventions could not be rejected.

 

Table 5           Summary of genetic programming technical analysis studies published between 1988 and 2004

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

1. Neely, Weller, 

    & Dittmar         

    (1997)

 

 

 

6 exchange rates: mark, yen, pound, Swiss franc, and two cross rates (mark/yen and pound/Swiss franc)

/ Daily

 

 

1975-77,

1978-80,

(1981-95)

 

 

100 trading rules generated by genetic programming during each in-sample period

 

Buy & hold

/ Optimized trading rules

 

In-sample periods: 0.1% per round-trip; out-of-sample period: 0.05%

 

Out-of-sample, genetic trading rules generated positive mean excess returns after transaction costs for every currency tested.  The mean excess return across all currencies was 2.9% per year, being higher than the B&H return (0.6%).  Since betas for these trading rule returns against various world market indices were negative, the excess returns did not seem to be compensation for bearing systematic risk.  In addition, the superior performance of trading rules could not be explained by standard statistical models such as a random walk, ARMA, and ARMA-GARCH.

 

2. Allen &

    Karjalainen

    (1999)

 

 

 

 

S&P 500 Index

/ Daily

 

1929-82

(1936-95)

 

100 trading rules generated by genetic programming during each in-sample period

 

Buy & hold

 

One-way transaction costs of 0.1, 0.25, and 0.5%

 

After considering reasonable one-way transaction costs of 0.25%, average excess returns of optimal trading rules were negative for 9 of 10 out-of-sample periods.  Even after transaction costs of 0.1%, average excess returns were negative for 6 out of the 10 periods.  In most periods, only a few trading rules indicated positive excess returns.  Overall, genetically formulated trading rules did not generate excess returns over the B&H strategy after transaction costs.

 

3. Fyfe, Marney,

    & Tarbert

    (1999)

 

 

U.K. Land Securities

/ Daily

 

1980-82,

1982-84

(1985-97)

 

The fittest trading rule generated by genetic programming during an in-sample period

 

Buy & hold

/ Optimized trading rules

 

1% per one-way transaction

 

Although an optimal trading rule performed well during the out-of-sample period, it appeared to have a similar structure to the B&H strategy.  When the optimal trading rule was applied to price series bootstrapped by three popular statistical models (a random walk, AR (1), AR (1)–ARCH (3)), only the AR (1) model explained about 40% of the original excess trading returns. 

 

4. Neely & Weller

    (1999)

 

 

 

 

 

4 cross exchange rates (mark/franc, mark/lira, mark/guilder, mark/pound)

/ Daily

 

1979-86

(1986-96)

 

 

100 trading rules generated by genetic programming, moving average (1/10, 1/50, 5/10, and 5/50 days), and filter (0.5, 1, 1.5, and 2%)

 

Buy & hold

/ Optimized trading rules

 

In-sample periods: 0.1% per round-trip; Out-of-sample period: 0.05%

 

 

During the out-of-sample period, annual mean excess returns averaged across 100 rules after transaction costs were positive for all four currencies, ranging 0.1% for the mark/guilder to 2.8% for the mark/pound.  In contrast, moving average rules and filter rules generated annual mean excess returns of -0.1% and -0.2% across all currencies, respectively.  There was no evidence that the excess returns to genetic trading rules were compensation for bearing systematic risk.

 

Table 5 continued.

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

5. Wang (2000)

 

 

 

S&P Index and S&P Index Futures

/ Daily

 

1984-97

(1987-98)

 

10 trading rules generated by genetic programming during each in-sample period

 

Buy & hold

/ Optimized trading rules

 

$0.50 per share + $25 per one-way transaction for spot index; $61 per round-trip for futures

 

For S&P futures, 36 out of 120 trading rules over the entire sample period outperformed the B&H strategy in terms of net returns.  However, the results varied from year-to-year.  Similar results were found when both S&P spot and futures markets were simultaneously considered for trading.  When risk-adjusted returns were assessed, 57 out of 120 rules beat the B&H strategy.  Although the performance of trading rules was still inconsistent over sample periods, more than 40% of the rules appeared to have some market-timing capability.   

 

6. Neely & 

    Weller (2001)

 

4 foreign exchange rates: mark, yen, pound, and Swiss franc

/ Daily

 

 

1975-80

(1981-92),

1987-92 (1993-98)

 

 

100 trading rules generated by genetic programming during each in-sample period

 

Buy & hold

/ Optimized trading rules

 

In-sample periods: 0.1% per round-trip; out-of-sample period

: 0.05%

 

Over the period 1981-92, intervention information from the Fed substantially improved the profitability of optimal trading rules for pound and Swiss franc.  For example, the median portfolio rule increased annual excess returns from 0.5% to 7.2% per year for the pound.  In contrast, over the 1993-98 period, intervention information decreased the profitability of trading rules for all currencies but the mark.  Thus, intervention activity did not seem to be a general source of profits for technical traders.  

 

7. Korczak &

    Roger (2002)

 

 

24 stocks of the CAC40 Index of the Paris Stock Exchange

/ Daily

 

Ten 261-day periods over 1/97-11/99

(Ten 7-day periods)

 

Trading rules generated by genetic programming during each in-sample period

 

Two buy & hold strategies

/Optimized trading rules

 

0.25% per one-way transaction

 

Out-of-sample results indicated that genetic trading rules outperformed both B&H strategies in 9 out of 10 cases.  Although newly generated trading rules performed well over time and relative to the old rules, all rules showed good and stable performance over the out-of-sample periods.  No trading rule consistently performed better than others. 

 

8. Ready (2002)

 

 

 

Dow Jones Industrial Average (DJIA)

/ Daily

 

1939-2000, 1957-62 (1963-86),

1981-86 (1987-00)

 

50 genetic-programming-based trading rules and 4 moving average rules from Brock et al. (1992)

 

Buy & hold,

Stock/bond weighted average

/ Optimized trading rules

 

0.13% per one-way transaction

 

Moving average rules generated positive excess returns after transaction costs for the period 1963-86, although they yielded negative excess returns for the period 1987-2000.  However, because moving average rules performed poorly from 1939-62, they were less likely to be chosen by traders at the beginning of 1963.  In fact, every genetic trading rule created over the period 1957-60 outperformed the moving average rules.  Similar results were found for the period 1987-2000.  Hence, Ready concluded that Brock et al.’s (1992) results for the period 1963-86 were spurious.   

 

Table 5 continued.

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

9. Neely (2003)

 

 

S&P 500 Index

/ Daily

 

1929-80

(1936-95)

 

10 trading rules generated by genetic programming during each in-sample period

 

Buy & hold

/ Optimized trading rules

 

0.25% per one-way transaction

 

During in-sample periods, genetic trading rules generated an about 5% annual mean excess return over the B&H strategy.  During out-of-sample periods, however, genetic trading rules generated negative mean excess returns over the B&H strategy.  The risk-adjusted performance based on several risk-adjusted return measures was inferior to that of the B&H strategy.  In addition, trading rules optimized by various risk-adjusted criteria also failed to outperform the B&H strategy.

 

10. Neely &

      Weller

      (2003)

 

4 foreign exchange rates: mark, yen, pound, and Swiss franc

/ Intra-daily

 

2/96-5/96 (6/96-12/96)

 

25 trading rules generated by genetic programming for each currency;

 

 

An linear forecasting model

/ Optimized trading rules

 

0, 0.01, 0.02 and 0.025% per one-way transaction

 

There was strong evidence of predictability in exchange rate series tested because genetically trained trading rules yielded annual returns of over 100% with zero transaction costs in 3 of the 4 cases.  However, under realistic trading hours and transaction costs (0.025%), genetic trading rules realized break-even transaction costs of less than 0.02% per one-way trade in all the exchange rates but the pound.  Moreover, genetic trading rules appeared to be inferior to the autoregressive linear forecasting model in most cases, although their performances were not much different. 

 

11. Roberts

      (2003)

 

 

CBOT corn, soybean, and wheat futures

/ Daily

 

1978-1998

(1980-1998)

 

The best of ten rules optimized during each in-sample period using genetic programming

 

Zero profits and buy & hold

 

$25 and $6.25 per contract per round-trip for in- and out-of-sample periods, respectively

 

Although genetically trained rules produced positive mean net returns only for wheat futures in out-of-sample tests, only trading rules that use the ratio of profit to maximum drawdown as a performance measure generated a statistically significant mean daily net profit of $0.93 per contract.  This was compared to the B&H profit of -$3.30 per contract.  For corn and soybean futures, however, genetic trading rules produced both negative mean returns and negative ratios of profit to maximum drawdown during the sample period.

 

 

Table 6           Summary of Reality Check technical analysis studies published between 1988 and 2004

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

1. Sullivan,

    Timmermann,

    & White (1999)

 

 

 

 

Dow Jones Industrial Average (DJIA), S&P 500 index futures

/ Daily

 

DJIA: 1897-1996,

1897-1986

(1987-96);

S&P 500 futures: 1984-96

 

Filter, moving average, support and resistance, channel breakout, on-balance volume average

 

Zero mean profits for mean return, a risk-free rate for the Sharpe ratio

/ Optimized trading rules

 

Not adjusted

 

During the 1897-96 period, the best rule in terms of mean return was a 5-day moving average that produced an annual mean return of 17.2% with a data snooping adjusted p-value of zero.  The corresponding break-even transaction cost was 0.27% per trade.  The best rule in terms of the Sharpe ratio generated a value of 0.82 with a Bootstrap Reality Check p-value of zero, while the B&H strategy generated a Sharpe ratio of 0.034.  However, during the 1987-96 period, the 5-day moving average rule earned a mean return of 2.8% per year with a nominal p-value of 0.32.  Moreover, in the S&P 500 futures market, the best rule generated a mean return of 9.4% per year with a Bootstrap Reality Check p-value of 0.90, implying that the return resulted from data snooping.

 

2. Qi & Wu (2002)

 

7 foreign exchange rates: mark, yen, pound, lira, French franc, Swiss franc, and Canadian dollar

/ Daily

 

 

1973-1998

 

Filter, moving average, support and resistance, and channel breakout

 

Buy & hold,

Zero mean profits

/

 

Adjusted

 

During the sample period, the best trading rules, which are mostly moving average rules and channel breakout rules, produced positive mean excess returns over the buy-and-hold benchmark across all currencies and had significant data snooping adjusted p-values for the Canadian dollar, the Italian lira, the French franc, the British pound, and the Japanese yen.  The mean excess returns were economically substantial (7.2% to 12.2%) for all the five currencies except for the Canadian dollar (3.6%), even after adjustment for transaction costs of 0.04% per one-way transaction.  In addition, the excess returns could not be explained by systematic risk.  Similar results were found for the Sharp ratio criterion, and the overall results appeared robust to incorporating transaction costs into the general trading model, changes in a vehicle currency, and changes in the smoothing parameter in the stationary bootstrap procedure. 

 

3. Sullivan,

    Timmermann,

    & White (2003)

 

 

 

Dow Jones Industrial Average (DJIA), S&P 500 index futures

/ Daily

 

DJIA: 1897-1998, 1987-96; S&P 500 futures: 1984-96

 

Technical trading systems from Sullivan et al. (1999) and calendar frequency trading rules from Sullivan et al. (2001)

 

Buy & hold

/ Optimized trading rules

 

Not adjusted

 

For the full sample period (1897-1998), the best of the combined universe of trading rules, a 2-day-on-balance volume strategy, generated a mean return of 17.1% on DJIA data with a data snooping adjusted p-value of zero, and outperformed the B&H strategy (a mean return of 4.8%).  For a recent period (1987-96), the best rule, a week-of-the-month strategy, produced a mean return of 17.3% slightly higher than the B&H return (13.6%), but the return was not statistically significant (p-value of 0.98).  Similar results were found for the S&P 500 futures data.  Although the best rule (a mean return of 10.7%) outperformed the benchmark (mean return of 8.0%) during the 1984-96 period, the data snooping adjusted p-value was 0.99. 

 

 

Table 7           Summary of chart pattern studies published between 1988 and 2004

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

1. Curcio,

    Goodhart,

    Guillaume,

    & Payne

    (1997)

 

3 foreign currencies: mark, yen, and pound

/ Intradaily (one hour frequency)

 

 

4/89-6/89,

1/94-6/94

 

 

Support and resistance, high-low, minimum of the support and low and maximum of the resistance and high, and max-min

 

Buy & hold

 

Bid-ask spreads

 

Across exchange rates tested, the results of the earlier sample period indicated that only 4 of 36 buy and sell rules yielded statistically significant positive returns after transaction costs.  Max-min rules showed even worse performance.  For the later period, 10 rules had positive returns but 14 rules produced significantly negative returns.  Max-min rules all realized negative returns.        

 

2. Caginalp &

    Laurent

    (1998)

 

 

All world equity closed end funds listed in Barron’s and all S&P 500 stocks

/ Daily

 

4/92-6/96,

1/92-6/96

 

Candlestick patterns

 

Average return

 

Commissions ($20 for several thousand shares) and the bid-ask spread (0.1-0.3%)

 

Candlestick reversal patterns appeared to have statistically significant short-term predictive power for price movements.  Each of the patterns generated substantial profits in comparison to an average gain for the same holding period.  For the S&P 500 stocks, down-to-up reversal patterns produced an average return of 0.9% during a two-day holding period (annually 309% of the initial investment).  The profit per trade ranged from 0.56%-0.76% even after adjustment for commissions and bid-ask spreads on a $100,000 trade, so that the initial investment was compounded into 202%-259% annually. 

 

3. Chang &

    Osler (1999)

 

 

 

 

6 spot currencies: yen, mark, pound, Canadian dollar, Swiss franc, and French franc

/ Daily

 

 

1973-94

 

Head-and-shoulders,

moving average (1/5, 1/20, 5/20, 5/50, and 20/50 days), and momentum (5-, 20-, and 50-day lags)

 

Buy & hold,

Equity yields

 

 

0.05% per round-trip

 

Head-and-shoulders rules earned substantial returns for the mark and yen but not for other currencies.  Profits for the mark and yen were around 13% and 19% per year, respectively, with being higher than the corresponding B&H returns or U.S.  equity yields.  These results were evident even after adjusting for transaction costs, risk, or interest differentials.  However, moving average rules and momentum rules appeared to have significant predictive power for all six currencies.  Moreover, they easily outperformed head-and-shoulders rules in terms of total profits and Sharpe ratios.

 

4. Guillaume

    (2000)

 

 

 

 

 

3 exchange rates: mark/dollar, yen/dollar, dollar/pound

/ Intra-daily

 

4/89-6/89,

1/94-6/94

 

4 trading range breakouts with a 0.1% band

 

Buy & hold

 

Bid-ask spreads

 

For the first sample period, several trading rules generated statistically significant net profits, particularly, in trending markets such as the yen/dollar market.  For the second period, however, none of the trading rules produced significant net profits, even in trending markets.  In general, support-resistance rules performed better than Max-Min rules used in Brock et al. (1992).

 

Table 7 continued.

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

5. Lo,

    Mamaysky,

    & Wang

    (2000)

 

 

Individual NYSE/AMEX and Nasdaq stocks

/ Daily

 

 

 

1962-96

 

Head-and-shoulders (H&S)and inverse H&S, broadening tops and bottoms (T&B), triangle T&B, rectangle T&B, and double T&B

 

Not considered

 

 

Not adjusted

 

Pattern-recognition algorithms were used to detect 10 chart patterns in price series smoothed by using non-parametric kernel regressions.  The results of goodness-of-fit and Kolmogorov-Smirnov tests indicated that, in many cases, return distributions conditioned on technical patterns were significantly different from unconditional return distributions, especially, for the Nasdaq stocks.  This suggests that technical patterns may provide some incremental information for stock investment, even if they may not be used to generate excess trading profits.

 

6. Osler (2000)

 

 

 

 

 

3 foreign exchange rates: mark, yen, and pound against U.S. dollar

/ Intra-daily

 

1/96-3/98

 

Support and resistance

 

 

Not considered

 

Not adjusted

 

“Bounce frequency of support and resistance levels for each currency published by six firms was compared to that of artificial support and resistance levels.  Results indicated that trends in intra-daily exchange rates were interrupted at the published support and resistance levels more frequently than at the artificial ones.  The results were consistent across all three exchange rates and all six firms, although the predictive power of the published support and resistance levels varied.  Moreover, the results were statistically significant and robust to alternative parameterizations. 

 

7. Leigh, Paz,

    & Purvis

    (2002)

 

 

 

The NYSE Composite Index

/ Daily

 

1980-99

 

Bull flag charting patterns

 

Buy & hold

 

Not adjusted

 

Across all parameter combinations considered, trading rule returns in excess of the B&H strategy were positive for all forecasting horizons (10, 20, 40, and 80 days).  Moreover, results of linear regression analyses indicated that trading rule parameters had predictive value for both price level and future price direction. 

 

8. Leigh,

    Modani,

    Purvis, &

    Roberts

    (2002)

 

The NYSE Composite Index

/ Daily

 

1980-99 (the first 500 trading days)

 

Two bull flag patterns with trading volume (a buy position is held for 100 days)

 

Buy & hold

/ Optimized parameters

 

Not adjusted

 

During the out-of-sample period, patterns outperformed the B&H strategy.  The first and the second bull flag patterns with trading volume generated statistically significant mean returns of 14.0% (with 55 buy signals) and 8.6% (with 132 buy signals) for 100-day holding period, respectively, while the B&H strategy profited 5.5%. 

 

9. Dawson & Steeley (2003)

 

 

225 individual FTSE100 and FTSE250 stocks

/ Daily

 

 

 

1986-2001

 

The same patterns as in Lo et al. (2000)

 

Buy & hold

 

 

Not adjusted

 

This study replicates Lo et al.’s (2000) procedure on UK data.  Results were similar to Lo et al.’s finding.  The results of goodness-of-fit and Kolmogorov-Smirnov tests indicated that return distributions conditioned on technical patterns were significantly different from the corresponding unconditional distributions.  However, across all technical patterns and sample periods, an average market adjusted return turned out to be negative.

 

Table 7 continued.

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

10. Lucke

      (2003)

 

Dollar, mark, pound, yen, and Swiss franc

/ Daily

 

1973-99

 

Head-and-shoulders

 

Not considered

 

Not adjusted

 

In general, head-and-shoulders rules failed to generate positive mean returns for all holding periods (1 to 15 days) except a one-day holding period.  In addition, it appeared that trading rule profits were not correlated with central bank intervention. 

 

11. Zhou &

Dong (2004)

 

 

 

 

 

 

 

1451 stocks listed on the NYSE, Amex, NASDAQ

/ Daily

 

1962-2000

 

Head-and-shoulders (HS)and inverse HS (HIS), broadening tops (BT) and bottoms (BB), triangle tops (TT) and bottoms (TB), rectangle tops (RT) and bottoms (RB)

 

Returns for a size- and momentum-matched control company

 

Not adjusted

 

To reflect the uncertainty of human perception and reasoning, fuzzy logic were incorporated into the definition of well-known technical patterns.  For all stocks tested, the HS, HIS, RT, and RB patterns generated significant cumulative abnormal returns (CARs) of around 3% for 120 days.  For stocks trading above $2.00, however, the significance of CARs dramatically reduced or disappeared.  The effect of small trading prices was more severe for NASDAQ stocks.  For the HS, IHS, and RB patterns, the fuzzy logic-based algorithm appeared to detect subtly different post-pattern performances between two portfolios with different pattern membership values.  The results for four subperiods indicated that for the RT pattern the post-pattern performances of two portfolios with different membership values were significantly different in the first three subperiods from 1962 through 1990.  This may imply that stock markets have been efficient after the early 1990s. 

 

Table 8           Summary of nonlinear technical analysis studies published between 1988 and 2004

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

1. Gençay

    (1998a)

 

Dow Jones Industrial Average (DJIA)

/ Daily

 

 

1963-88

(Last 250 prices for each of 6 sub-samples)

 

Trading rules based on a feedforwad network model

 

Buy & hold

/ Optimized models

 

$600 per round-trip for the contract value of 1,000,000

 

Trading signals as a function of past returns were generated by a feedforward network, which is a class of artificial neural networks.  Across subperiods, net returns of technical trading rule (7% to 35%) dominated those of the B&H strategy (-20% to 17%).  Sharpe ratio tests indicated similar results.  Correct sign predictions for the recommended positions ranged from 57% to 61% for all subperiods.

 

2. Gençay

    (1998b)

 

 

Dow Jones Industrial Average (DJIA)

/ Daily

 

 

 

 

1897-1988 (10 most recent prices for each of 22 sub-samples)

 

 

Trading rules based on a feedforwad network model

 

An OLS model with lagged returns as regressors

/ Optimized models

 

Not adjusted

 

In terms of forecast improvement measured by the mean square prediction error (MSPE), non-linear models (feedforward network models) using past buy-sell signals from moving average rules (1/50 and 1/200) as regressors outperformed linear specifications such as the OLS, GARCH-M (1,1), and a feedforward network regression with past returns.  For 14 of 22 subperiods, the nonlinear models generated at least 10% forecast improvement over the benchmark model.  The model with a 1/50 moving average rule provided more accurate out-of-sample predictions relative to one with a 1/200 rule.

 

3. Gençay &

Stengos (1998)

 

 

 

 

 

 

 

 

Dow Jones Industrial Average (DJIA)

/ Daily

 

1963-88

(Last 1/3 of the data set for each of 6 sub-samples)

 

Trading rules based on a feedforwad network model

 

An OLS model with lagged returns as regressors

/ Optimized models

 

Not adjusted

 

Overall non-linear models (feedforward network models) outperformed linear models (OLS and GARCH-M (1,1)) in terms of MSPEs and sign predictions.  The non-linear models with lagged returns generated an average of 2.5% forecast improvement over the benchmark model with lagged returns.  This prediction power improved as large as 9.0% for the non-linear models in which past buy-sell signals of a moving average rule (1/200) were used as regressors.  In particular, when the non-linear model included a 10-day volume average indicator as an additional regressor, it produced an average of 12% forecast gain over the beanchmark and provided much higher correct sign predictions (an average of 62%) than other models.    

 

4. Gençay   

    (1999)

 

 

5 spot exchange rates: pound, mark, yen, France franc, and Swiss franc

/ Daily

 

 

1973-92

(Last 1/3 of the data set)

 

Trading rules based on a feedforwad network model and the nearest neighbor regression

 

Random walk and GARCH (1,1) models

/ Optimized models

 

Not adjusted

 

Nonlinear models such as the nearest neighbors and the feedforward network regressions with past buy-sell signals from moving average rules (1/50 and 1/200) outperformed a random walk and a GARCH (1,1) model in terms of sign predictions and mean square prediction errors.  For example, average correct sign prediction of the nearest neighbors model was 62% for the five currencies.  Models with a 1/50 moving average rule provided more accurate predictions over models with a 1/200 rule. 

 

Table 8 continued.

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

5. Fernández-

    Rodríguez,

    González-

    Martel, &

    Sosvilla-

    Rivero (2000)

 

The General Index of the Madrid Stock Market

/ Daily

 

1966-97

(10/91-10/92, 7/94-7/95, 10/96-10/97)

 

A trading rule based on a feedforwad network model

 

Buy & hold

 

 

Not adjusted

 

 

 

 

 

In terms of gross returns, a trading rule based on a feedforwad network model dominated the B&H strategy for two subperiods, while the opposite was true for most recent subperiods in which there exists upwards trend.  Correct sign predictions for the recommended positions ranged from 54-58%, indicating better performance than a random walk forecast.  

 

6. Sosvilla-

    Rivero,

    Andrada-

    Félix,

    & Fernández-

    Rodríguez

    (2002)

 

Mark and yen

/ Daily

 

1982-96

 

A trading rule based on the nearest neighbor regression

 

Buy & hold

/ Optimized models

 

0.05% per round-trip

 

Trading rule generated net returns of 35% and 28% for the mark and yen, respectively, and outperformed B&H strategies that yielded net returns of -1.4% and -0.4%, respectively.  Correct sign predictions for recommended positions were 53% and 52% for the mark and yen, respectively, beating a random walk directional forecast.  However, when excluding days of US intervention, net returns from the trading strategy substantially decreased    (-10% and -28% for the mark and yen, respectively) and were less than the B&H returns in both cases. 

 

7. Fernández-

    Rodríguez,

    Sosvilla-

    Rivero, & 

    Andrada-Félix

    (2003)

 

9 exchange rates in the European Monetary System (EMS)

/ Daily

 

1978-94,

 

 

Trading rules based on the nearest neighbor (NN) and the simultaneous NN regressions and moving averages (1/50, 1/150, 1/200, 5/50, and 5/200 days)

 

 

Not considered

/ Optimized models

 

0.05% per round-trip

 

For most exchange rates, annual mean returns from nonlinear trading rules based on the nearest neighbor or the simultaneous nearest neighbor regressions were superior to those of moving average rules.  The nonlinear trading rules also generated statistically significant annual net returns of 1.5%-20.1% for the Danish krona, French franc, Dutch guilder, and Italian lira.  Similar results were found for the Sharp ratio criterion.  The nonlinear trading strategies generated the highest Sharpe ratios in 8 out of the 9 cases.  

Table 9           Summary of other technical analysis studies published between 1988 and 2004

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

1. Pruitt &

    White (1988)

 

204 stocks from the CRSP at the University of Chicago

/ Daily

 

1976-85

 

CRISMA (combination system of Cumulative volume, RelatIve Strength, and Moving Average)

 

Buy & hold

 

0 to 2% per round-trip

 

After 2% transaction costs and across various return-generating models, the CRISMA system yielded annual excess returns ranging from 6.1% to 15.1% and beat the B&H or market index strategy.  The system also generated a much greater percentage of profitable trading successes after transaction costs than would be expected by chance.  

 

2. Schulmeister

    (1988)

 

Mark

/ Daily

 

1973-88

 

Moving average,

momentum, point & figure, combination of moving average & momentum

 

Buy & hold

 

0.04% per one-way transaction

 

All trading rules considered produced substantial annual returns up to 16%.  The combination system performed best.  The probability of an overall loss appeared to be less than 0.005% when one of the trading rules was followed blindly during the 1973-86 period.

 

3. Sweeney

    (1988)

 

 

 

14 Dow-Jones Industrial stocks

/ Daily

 

1956-62

(1970-82)

 

 

0.5% filter rule

 

Buy & hold

 

From 0.05% to 0.2% per one-way transaction

 

During the 1970-82 period, for 11 of 14 stocks that had earned profits before commissions in Fama and Blume’s (1966) study, a 0.5% filter rule produced statistically significant annual mean returns after adjustment for transaction costs of 0.1%.  For an equally weighted portfolio of 14 stocks, the filter rule generated a mean net return of 10.3% per year.  Portfolio returns appeared to be robust across several subsamples but were quite sensitive to transaction costs. 

 

4. Taylor (1988)

 

Treasury bond futures from CBOT

/ Daily

 

1978-87

 

A statistical price-trend model based on ARMA(1,1)

 

Buy & hold

 

0.2% per round-trip

 

All four trading rules generated positive average excess returns ranging from 4.4% to 6.8% per year and were superior to the B&H strategy.  However, t-test results indicated that none of the returns was significantly different from zero at the 5% level.  In addition, the B&H strategy performed better than each trading rule from 1982-87. 

 

5. Pruitt &

    White (1989)

 

In-the-money call options written on the 171 stocks

/ Daily

 

1976-85

 

CRISMA

 

Not considered

 

Maximum 1988 retail transaction costs

 

After transaction costs, the CRISMA system generated a mean return of 12.1% per round trip.  In fact, 71.3% of the 171 transactions were profitable after adjustment for transaction costs.  The binomial proportionality test statistics showed that the trading profitability could not be achieved by chance. 

 

6. Neftci (1991)

 

 

Dow-Jones Industrials

/ Monthly

 

 

1792-1976

 

Moving average

(150 days)

 

Not considered

 

Not adjusted

 

This study showed that moving average rules were one of the few statistically well-defined procedures.  Trading signals of a 150-day moving average rule were incorporated into a dummy variable in an autoregression equation.  F-test results on the variable were insignificant for 1795-1910 but highly significant for 1911-76, indicating some predictive power of the moving average rule.

 

Table 9 continued.

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

7. Corrado &

    Lee (1992)

 

 

 

120 stocks from the Dow Jones and S&P 500 Index

/ Daily

 

 

1963-89

 

0.5% own-stock filter,

0.25% S&P 500 Index filter,

0.5% other-stock filter

 

Buy& hold

 

0.04% per one-way transaction

 

The own-stock filter rule generated an equally-weighted mean portfolio return of 30.8% per year during the sample period, while the B&H strategy yielded a mean portfolio return of 11.3% per year.  This difference between the returns made an annual gross margin of 6.4% over the B&H strategy after transaction costs.  

 

8. Pruitt, Tse, &

    White (1992)

 

148 stocks and

in-the-money call options written on the 126 target stocks

/ Daily

 

1986-90

 

CRISMA (combination system of Cumulative volume, RelatIve Strength, and Moving Average)

 

Buy& hold

 

Security: 0-2% per round-trip; Option: $60 per round-trip

 

For stocks, the CRISMA system generated annualized excess returns of between 1.0% and 5.2% after transaction costs of 2% and outperformed the B&H or market index strategy.  For options, the system generated highly significant returns of 11.0% per option trade after transaction costs, with 63.5% of all trades being profitable.  

 

9. Wong (1995)

 

 

 

 

 

Hang Seng Index (HSI)

/ Daily

 

1969-1990,

5 subperiods

 

 

Moving average

(10, 20, and 50 days)

 

Buy & hold

 

Not adjusted

 

In general, moving average rules performed well.  In particular, an MA10 (a 10-day moving average) bullish signal, an MA20 bullish signal, and an MA50 bearish signal generated statistically significant excess returns over the B&H strategy.  It appeared that for buy (sell) signals, prices declines (rises) slowly in the early pre-event period and rises (declines) sharply in the late pre-event period.  Prices continued to rise (declines) slowly in the post-event period for buy (sell) signals.

 

10. Cheung &

      Wong

      (1997)

 

Yen, Singapore dollar, Malaysian ringgit, and Taiwan dollar

/ Daily

 

1986-95

 

Filter

(0.5, 1, and 1.5%)

 

Buy & hold

 

1/8 of 1% of asset value per round-trip

 

 

When transaction costs and risk were adjusted, filter rules generated superior excess returns over the B&H strategy only for the Taiwan dollar.  Filter rules were inferior to the B&H strategy in the cases of the yen and Singapore dollar.  Both filter rule and B&H strategies failed to generate significant excess returns on the Malaysian ringgit. 

 

11. Irwin,

      Zulauf,

     Gerlow, &

     Tinker (1997)

 

Futures contracts for soybean, soybean meal, and soybean oil

/ Daily and monthly

 

1974-83

(1984-88)

 

Channel (40 days),

ARIMA(2,0,0) for soybean and ARIMA(1,0,1) for soybean mean and oil

 

Zero mean profits

 

Not adjusted

 

During the out-of-sample period, the channel system generated statistically significant mean returns ranging 5.1%-26.6% for all markets.  The ARIMA models also produced statistically significantly positive returns (16.5%) for soybean meal, but significantly negative returns (-13.5%) for soybeans.  For every market, the channel system beat the ARIMA models.

 

Table 9 continued.

 

 

                Criteria:

 

Study

 

Markets considered

/ Frequency of data

In-sample period (Out-of-sample period)

 

Technical trading systems

 

Benchmark strategies / Optimization

 

Transaction

costs

 

Conclusion

 

 

12. Neely (1997)

 

 

 

 

4 foreign currencies: mark, yen, pound, and Swiss franc

/ Daily

 

1974-97

 

Filter (0.5, 1, 1.5, 2, 2.5, and 3%) and

moving average (1/10, 1/50, 5/10, and 5/50 days)

 

Buy & hold the S&P 500 index

 

0.05% per round-trip

 

Technical trading rules showed positive net returns in 38 of the 40 cases.  In general, moving average rules performed slightly better than filter rules.  Moreover, the trading profits were not likely to be compensation for bearing risk.  For example, for the mark, every moving average rule beat the B&H strategy of the S&P 500 Index in terms of the Sharpe ratio.  The CAPM betas from the trading rules also generally indicated negative correlation with the S&P 500 monthly returns.  

 

13. Goldbaum

      (1999)

 

 

 

 

 

U.S. T-Bills, a value-weighted market portfolio of all the NYSE and AMEX securities from the CRSP, and IBM stock

/ Daily

 

1962-89

 

Moving average

(1/50, 1/200, 5/50, and 5/200 days with 0 and 1% bands)

 

T-Bill returns

 

Not adjusted

 

As a performance measure, the price error between assets was estimated using the nonparametric stochastic discount factor (SDF), which was either conditioned or unconditioned on public information (e.g.  term structure).  For the market portfolio returns, moving average rules generally had unconditional estimates that were significantly positive or close to zero and conditional estimates that were negative or close to zero, implying a negative performance of the trading rules to an informed trader.  For IBM stock returns, however, the conditional estimates on the term structure were significantly different from zero. 

 

14. Marsh

      (2000)

 

 

3 IMM currency futures: mark, yen, and pound sterling

/ Daily

 

1980-96,

1980-85

(1986-90),

1980-90

(1991-95)

 

Markov models and

moving average rules

(1/5, 5/20, and 1/200 days)

 

Not considered

 

0.025% and 0.04% per one-way transaction

 

Before transaction costs, all moving average rules tested yielded positive returns for both 1981-85 and 1986-90, but the rules generated positive returns only in 3 out of 9 cases for 1991-95.  For out-of-sample periods, Markov models also generated positive returns in 2 out of 6 cases.  Augmented Markov models, in which interest differentials were included, produced substantially positive returns for all 3 currency futures during 1986-90 but only for the yen during 1991-95.   

 

15. Dewachter

      (2001)

 

 

 

 

 

 

4 foreign exchange rates: mark, yen, pound, and franc

/ Weekly

 

1973-97

 

Moving average (1/30) with a 5-day holding period,

Markov model and its ARMA (1,1) representation as the class of Taylor’s price-trend models

 

Not considered

 

Not adjusted

 

Across exchange rates, the moving average rule produced a statistically significant average return of about 6% per year and the correct sign prediction of about 55%.  The extended Markov switching model and the ARMA (1,1) representation of the Markov switching model showed even better performance in terms of profits and sign prediction.  The results of Monte Carlo simulations indicated that the Markov model could replicate the observed profitability of the moving average rule. 

 

16. Wong,

      Manzur, &

      Chew (2003)

 

 

Singapore Straits Times Industrial Index (STII)

/ Daily

 

1974-1994,

Three 7-year subperiods

 

Moving averages and

relative strength index (RSI)

 

Not considered

 

Not adjusted

 

In general, every trading system tested produced statistically significant returns over all three subperiods and a whole period.  Single moving average rules generated the best results, followed by dual moving average crossover rules and relative strength index rules. 

 


Table 10        The profitability of technical trading strategies in modern studies (1988-2004)

 

 

Studies

The number of studies

Net profit range

 (Out-of-sample period)

 

Comments

Positive

Mixed

Negative

A. Stock markets

 

 

 

 

1.1%a

(1968-88)

• For the Dow Jones Industrial Average (DJIA) data, which was most frequently tested in the literature, results varied considerably depending on the testing procedure adopted.  In general, technical trading strategies were profitable until the late 1980s.  However, technical trading strategies were no longer economically profitable thereafter.

• Overall, variable-moving average rules showed a quite reliable performance for the stock market over time.

• For several non-US stock markets (e.g., Mexico, Taiwan, and Thailand), moving average rules generated large annual net profits of 10% to 30% until the mid-1990s. 

Standard

1

0

3

Model-based Bootstrap

7

2

3

Genetic programming

2

1

3

Reality Check

0

1

1

Chart patterns

5

0

1

Nonlinear

3

0

1

Others

8

1

0

Sub-total

24

5

12

 

B. Currency markets

 

 

 

 

 

5%-10%

(1976-91)

 

• Many studies investigated major foreign currency futures contracts traded on the CME, i.e., the Deutsche mark, Japanese yen, British pound, and Swiss franc. 

 

• For major currencies, a wide variety of technical trading strategies, such as moving average, channel, filter, and genetically formulated trading rules, consistently generated economic profits until the early 1990s. 

 

• Several recent studies confirmed the result, but also reported that technical trading profits have declined or disappeared since the early 1990s, except for the yen market. 

Standard

7

3

3

Model-based bootstrap

6

0

1

Genetic programming

3

0

1

Reality Check

1

0

0

Chart patterns

2

0

3

Nonlinear

3

0

0

Others

3

1

1

      Sub-total

25

4

9

 

C. Futures markets

 

 

 

 

 

4%-6%

(1976-86)

 

 

• Technical trading strategies generated economic profits in futures markets from the late 1970s through the mid-1980s.  In particular, technical trading strategies were consistently profitable in most currency futures markets, while they appeared to be unprofitable in livestock futures markets.   

 

• Channel rules and moving average rules were the most consistent profitable strategies. 

 

• After the mid-1980s, the profitability of technical trading strategies for overall futures markets were not investigated comprehensively yet.

Standard

5

0

1

Model-based bootstrap

1

0

1

Genetic programming

0

1

0

Others

1

0

1

      Sub-total

7

1

3

Total

58

10

24

 

 

a This is a break-even one-way transaction cost of a 5/200 moving average rule, which was optimized by using the DJIA data from 1897 to 1968 (Taylor 2000).