Technical analysis uses recurring chart patterns, such as head-and-shoulders and double tops, to predict future price movements from historical price data. Lo, Mamaysky, and Wang (2000) used kernel regression in Foundations of Technical Analysis: Computational Algorithms, Statistical Inference, and Empirical Implementation to detect 10 chart patterns. They applied this method to CRSP data on 350 NYSE/AMEX stocks and 350 Nasdaq stocks from 1962 to 1996. All 10 technical patterns tested were statistically significant for Nasdaq stocks at the 5 percent level, versus only 5 of 10 for NYSE/AMEX stocks.
What the Study Found
Head-and-shoulders patterns occurred 1,611 times in actual NYSE/AMEX data versus only 577 times in simulated geometric Brownian motion. Broadening top patterns occurred only 725 times in actual data, compared with 1,227 times in the simulated random-walk benchmark. Broadening tops with increasing volume trend occurred 409 times, compared with 143 occurrences with decreasing volume trend. A chi-square goodness-of-fit test found significant differences for 7 of the 10 patterns in NYSE/AMEX stocks. The weakest result was a 21.2 percent p-value for triangle top patterns.
Methodology
The study draws on daily stock return data from the Center for Research in Securities Prices (CRSP) database. The sample consists of 350 NYSE/AMEX stocks and 350 Nasdaq stocks, with 50 stocks drawn from each of five market-capitalization quintiles in every five-year subperiod. Returns are measured from 1962 to 1996, split into seven five-year subperiods: 1962 to 1966, 1967 to 1971, and so on. The analysis controls for market-capitalization quintile, exchange listing, and volume trend, comparing periods of increasing versus decreasing average share turnover.
Key Statistics
| Metric | Finding | Context |
|---|---|---|
| Head-and-shoulders frequency | 1,611 actual vs. 577 simulated GBM | NYSE/AMEX stocks, 1962-1996 |
| Kolmogorov-Smirnov significant patterns | 5 of 10 patterns significant (p = 0.000-0.021) | NYSE/AMEX stocks, 1962-1996 |
| Kolmogorov-Smirnov significant patterns | All 10 patterns significant at 5% level | Nasdaq stocks, 1962-1996 |
| Goodness-of-fit statistic Q | Q = Σ(nⱼ - 0.10n)² / 0.10n, distributed χ²(9) | Tests whether conditional returns are uniform across deciles of unconditional returns |
| Kernel regression bandwidth | 0.3 × h* (cross-validation bandwidth) | Used in rolling-window pattern detection over 38-day windows |
Why This Matters
The findings suggest that some classic chart patterns capture information not yet reflected in prices, particularly for less liquid Nasdaq stocks. Traders and quant researchers can treat this as evidence that automated, rule-based pattern detection is a viable alternative to subjective visual chart reading. Finding that a pattern is statistically informative does not mean it produces profitable trading strategies once transaction costs are considered. The weaker results for NYSE/AMEX stocks suggest any informational edge from technical patterns may already be partly arbitraged away in more liquid, heavily traded markets.