Scientific Technical Analysis for Traders

Evidence based technical analysis (EBTA) is dedicated to the proposition that technical analysis should be approached in a scientific manner. This implies several things. First, it is restricted to objective methods that can be simulated on historical data. Second, the historical performance statistics produced by such back-testing are then evaluated in a statistically rigorous fashion. In other words, profitable past performance is not taken at face value but rather evaluated in light of the possibility that back-test profits can occur by sheer luck. The problem of lucky performance is especially pronounced when many methods are back-tested and a best method is selected. This activity is called data mining. Though data mining is a promising approach for finding predictive patterns in data produced by largely random complex processes such as financial markets, its findings are upwardly biased. This is the data mining bias. Thus, the profitability of methods discovered by data mining must be evaluated with specialized statistical tests designed to cope with the data mining bias. EBTA employs such methods. EBTA rejects all subjective, interpretive methods of Technical Analysis as worse than wrong, because they are untestable. Thus classical chart patterns, Fibonacci based analysis, Elliott Waves and a host of other ill defined methods are rejected by EBTA. Yet there are numerous practitioners who believe strongly that these methods are not only real but effective. How can this be? Here, EBTA relies on the findings of cognitive psychology to explain how erroneous beliefs arise and thrive despite the lack of valid evidence or even in the face of contrary evidence. Cognitive psychologists have identified various illusions and biases, such as the confirmation bias, illusory correlations, hindsight bias, etc. that explain these erroneous beliefs. Thus EBTA relies on computerized methods for identifying patterns, and combining evidence into useful trading signals. Due to recent advances in computing and data mining algorithms it becomes possible for the modern technical analyst to amplify their research efforts and find the real gold. In other words, EBTA advocates a synergistic partnership between technical analysts and data mining computers to expand the valid base of knowledge called technical analysis. The union of humans and intelligent machines makes sense because the two entities have different but complimentary information processing abilities. Whereas human intelligence has a limited ability to engage in complex configural reasoning, which is required to identify valid predictive variables and combine them into a mathematical function, it can pose questions and proposed candidate variables. Whereas computer intelligence is ill equipped to pose questions and propose variables it has enormous capacities to identify relevant predictors and derive optimal combining functions. However, this new approach to technical analysis will require that human technicians abandon some tasks they now do and learn a new set of analytical skills. While they will no longer try to subjectively evaluate complex information patterns, they will need to learn about the kinds of data transformations that produce variables that are most digestible to data mining computers. They will also need to learn which data mining approaches are most viable and which types of problems are most amenable to data mining.

David Aronson is the author of “Evidence Based Technical Analysis” (John Wiley & Son’s 2006).

© Content Copyright 2006-2009 David Aronson. All rights reserved. Web Design & Graphics by eyeDisign.

Scientific Technical Analysis for Traders

The Scientific Method in Technical Analysis

At the heart of EBTA is the application of the scientific method to the study of financial markets. This requires moving beyond anecdote and intuition to formulate explicit, testable hypotheses about market behavior. A hypothesis in this context is a proposed rule or pattern, such as "a moving average crossover generates signals that outperform a buy-and-hold benchmark after transaction costs." The hypothesis must be specified precisely enough to be coded into an algorithm and simulated on historical data. Only then can its validity be assessed objectively.

Central to this process is the avoidance of common pitfalls that plague informal back-testing. Look-ahead bias occurs when a model uses information that would not have been available at the time of the trade, such as using the closing price to decide an entry at that same close. Survivorship bias arises when only currently existing securities are included in a historical dataset, omitting those that have been delisted or gone bankrupt. EBTA practitioners rigorously scrub data and verify execution assumptions to ensure that back-tested results are achievable in real time. Without such care, even a method with no true predictive power can appear remarkably profitable.

Moreover, EBTA insists on out-of-sample testing. A model that is optimized on one period of data may simply have memorized noise rather than captured a genuine market inefficiency. By reserving a portion of the data for validation—never used in the model selection process—the analyst can obtain an unbiased estimate of the method's likely future performance. This division between in-sample and out-of-sample data is a hallmark of serious quantitative research and a core discipline of evidence based technical analysis.

The Scientific Method in Technical Analysis

Back-Testing and Performance Evaluation

Back-testing is the engine of EBTA, but it must be conducted with an eye toward realism. The analyst must specify a complete trading system: entry and exit rules, position sizing, and the treatment of transaction costs, commissions, and slippage. Ignoring these frictions can turn a losing strategy into a seemingly winning one. For example, a high-frequency signal that trades daily may generate many small edges that are entirely consumed by costs. Only after accounting for such expenses can performance be evaluated meaningfully.

Performance is assessed through a suite of statistics, not just raw return. The Sharpe ratio measures return per unit of risk, while maximum drawdown captures the worst peak-to-trough decline. Win rate and profit factor are also common, but these numbers are only meaningful in the context of statistical significance. A method with a high Sharpe ratio but few trades may have achieved that result by luck. EBTA therefore applies inference tests—such as the bootstrap or Monte Carlo permutation—to determine whether the observed performance is unlikely to have arisen by chance alone.

A common exercise is to compare the tested system against a benchmark or a null model, such as a random walk. If the system's returns cannot be distinguished from what might be achieved by random trading on the same data, then it lacks genuine predictive power. EBTA demands that any claimed market inefficiency pass this statistical gauntlet before it is considered worthy of real capital.

Back-Testing and Performance Evaluation

The Data Mining Bias and Its Mitigation

Data mining bias, also known as the multiple testing problem, is a central concern in EBTA. When an analyst tests thousands of trading rules on the same dataset, some will appear profitable purely by chance, even if no true pattern exists. Selecting the best-performing rule from such a search yields an upwardly biased estimate of its future performance. This is especially problematic in finance, where large datasets and powerful computers make it easy to generate many candidate systems.

To counter this, EBTA adopts specialized statistical procedures. White's Reality Check and Hansen's Superior Predictive Ability (SPA) test are designed to adjust for the data mining bias when comparing a set of trading rules against a benchmark. These methods use bootstrap or simulation techniques to construct the distribution of the best performance under the null hypothesis of no predictive ability. Only if the observed best rule outperforms this distribution is it deemed statistically significant.

Additionally, EBTA advocates a disciplined model development process: keep a strict separation between the data used for exploration and the data used for final confirmation. Walk-forward analysis, where the model is periodically re-optimized on a rolling basis and then applied to the next out-of-sample segment, provides a more realistic assessment of how a strategy might perform in live trading. These techniques help ensure that the "gold" discovered by data mining is genuine and not fool's gold.

The Data Mining Bias and Its Mitigation

The Human-Computer Partnership in EBTA

The synergy between human intelligence and machine computation is pivotal. Humans excel at generating hypotheses, understanding market context, and identifying potentially informative variables. For instance, a technician might propose that volatility compression precedes major breakouts, or that certain economic data releases create predictable short-term reactions. The computer then tests these hypotheses across vast datasets, searching for confirming evidence or refutation.

To facilitate this, the analyst must transform raw price and volume data into features that are digestible to data mining algorithms. Common transformations include:

  • Technical indicators (moving averages, RSI, MACD)
  • Volatility measures (historical volatility, ATR)
  • Pattern recognition outputs (e.g., breakout flags, pivot points)
  • Intermarket data (correlations, relative strength)
  • Fundamental and sentiment data (if available)

The computer then applies machine learning methods—such as neural networks, support vector machines, or genetic algorithms—to combine these features into predictive models. However, EBTA is cautious about overfitting. The more complex the model, the more data it requires and the greater the risk of fitting noise. Thus, the human role includes selecting appropriate model complexity and validating results with out-of-sample tests. This partnership leverages the strengths of both: human creativity and machine discipline.

The Human-Computer Partnership in EBTA

Practical Implementation Challenges

Implementing EBTA in practice is not without obstacles. Data quality is paramount; errors in price data, missing observations, or incorrect adjustments for splits and dividends can distort back-tests. Survivorship bias, mentioned earlier, must be addressed by using point-in-time databases that include delisted securities. Furthermore, markets are non-stationary: a pattern that worked in the past may decay as it becomes widely known or as market structure changes. EBTA therefore requires continuous monitoring and adaptation of strategies.

Another challenge is the temptation to over-optimize. With enough parameters, a model can be made to fit any historical dataset almost perfectly, yet perform poorly out-of-sample. EBTA practitioners must resist this urge by limiting the number of free parameters, using parsimonious models, and relying on cross-validation. The goal is to find robust patterns that generalize, not to produce a back-test curve that looks perfect.

Finally, the psychological demands of systematic trading are significant. Even with a statistically validated edge, drawdowns and periods of underperformance can test the trader's discipline. EBTA encourages a rule-based approach that minimizes discretionary interference, since human emotions are often the enemy of consistent execution. For more on the origins and philosophy of this approach, see About Dave Aronson.

Practical Implementation Challenges

The Future of EBTA

As computational power continues to grow and machine learning techniques advance, the potential for evidence based technical analysis expands. The explosion of alternative data—such as social media sentiment, satellite imagery, and high-frequency order flow—offers new variables that might contain predictive information. However, the data mining bias becomes ever more severe as the search space grows, making rigorous statistical validation more critical than ever.

EBTA is poised to play a central role in this evolution. By combining the scientific method with powerful computing, it provides a framework for separating true market inefficiencies from random noise. The future may see even tighter integration between human analysts and intelligent machines, with automated hypothesis generation and real-time model updating. Yet the core principles—objectivity, testability, and statistical rigor—will remain unchanged. As David Aronson has argued, this is the only path to reliable knowledge in technical analysis.

The Future of EBTA

Evidence Based Technical Analysis