Three claims with different burdens of proof
Lo, Mamaysky and Wang replace subjective pattern recognition with a formal algorithm. They study historical US equities and find incremental information in some technical features in their sample. That is stronger than a chart anecdote, but conditional informativeness is not guaranteed profitability for every implementation and does not validate every popular indicator.[1]
A narrow research question might ask whether subsequent returns differ after a pre-defined event from a comparable control group. Specify assets, frequency, window, calculation time and result horizon. If a rule requires a daily closing price, do not automatically count a trade at that same price before it became known. Information availability is part of the hypothesis.
Freeze the rule before looking ahead
Create an immutable version of the rule before testing held-out data. Include missing-data treatment, corporate actions, trading halts and absence of a counterparty. If parameters change after viewing the held-out sample, that sample has become training data. The next independent test needs a new period or pre-specified group, not merely a new name for the same table.
Preserve the number of parameter families tested, window changes, excluded periods and selection criteria. One attractive result among many trials may be selection. A beautiful chart alone cannot establish persuasiveness, for indicators or fundamental ratios.
An original small-effect price example
Suppose a fictional rule produces an average price change of 0.12% per completed cycle. Commission and half the spread on two operations together cost 0.10%, and additional slippage costs 0.05%. After those stated costs, the result is −0.03%, although the gross-effect chart looked positive. This is arithmetic, not a measurement of a real instrument or a current broker tariff.
Even a positive remainder does not finish the test. Examine loss distribution, dependence between observations, the ability to obtain the required quantity and short availability where relevant. Events on the same day of a common shock are not mechanically independent repetitions. A closing price may poorly describe an achievable execution price for illiquid securities.
What remains off camera
Test robustness under small changes to the window and start date, but do not turn that into another search for a maximum. A narrow parameter combination without a clear mechanism is suspicious. Failure to find a stable effect does not prove that prices are perfect; efficiency studies also show limits on test power. “The rule was not confirmed” and “all market estimates are correct” remain logically different.[2]
Useful technical work can still measure a series consistently, compare intervals and expose data problems without promising trades. A moving average is a transparent filter even if a crossover has not been validated as a signal; pattern testing shows how to turn visual intuition into a reproducible hypothesis.
Data suitability and execution log
For a chosen frequency, define minimum coverage before viewing results: for example, at least 95 of 100 expected daily closes. Never replace an unknown close with the future; record no execution for a halted instrument. In a separate specification, median daily turnover over the previous 20 sessions must exceed 2 million units and an order must not exceed 1% of turnover, giving a calculated 20,000 limit at 2 million. This is a stress-test subject, not a universal safe-stock standard.
Exclude a security from a test when indicator history is insufficient, a split leaves an unresolved scale error or the availability date of a key field cannot be reconstructed. Record the reason and do not use excluded results to choose the filter.
Use execution-log rows such as event detected, order permitted, executed, exit requested and closed. A fictional event is visible at the 18:00 close and submission is permitted in the next session. A signal price of 100, next achievable purchase price of 101 and sale price of 102 produce a 2% signal-to-sale change but a gross realized result of 102/101−1≈0.9901%. With 0.3% cycle costs on initial capital, net result is approximately 0.6901%. These figures are illustrative, not a current tariff.
Test the process without turning it into a trading course
First run the log on hand-checkable artificial series: a signal on the last day with no next session, a halt, a price gap and partial execution. Then use a shadow test without real money, freezing event time before future quotes. This tests discipline and calculation, not predictive value; the Lo study’s empirical findings are not inherited by this log.[RTA]
A completion criterion is that another researcher using the same data version obtains the same list of permissions, signals, unfilled orders and costs. Independent evaluation, risk and regime change still remain.
Three “axioms” as testable hypotheses
Replace “price reflects everything” with a limited hypothesis: a selected public feature known at time t does not improve a pre-defined forecast versus a control after risk and costs. One feature’s lack of improvement does not prove a claim about all information.
“Prices move directionally” requires a horizon and variable. One protocol could compare the sign of the last five non-overlapping returns with the sign of the next five, preserving all cases and a control with the same share of rising periods. A visual trend or positive average change does not prove predictable signs.
“History repeats” becomes a transfer hypothesis: test the same frozen classification rule in the next period on comparable data, with metric, material effect size, precision and possible regime-change indicators fixed beforehand. Lo et al. support separating algorithmic definition from conditional distribution; they do not prove these general axioms.
Sources and limits of review
[1] [Foundations of Technical Analysis](https://www.nber.org/papers/w7613): formalization of chart patterns and conditional historical informativeness, not proof of profit. [2] [Do We Really Know That Financial Markets Are Efficient?](https://www.nber.org/papers/w0994): limits of empirical efficiency tests over long horizons. [RTA] Lo, Mamaysky and Wang, “Foundations of Technical Analysis,” source rm-technical-lo; the suitability filters and execution log above are an independent educational protocol.