Strategy testing protocol — backtest and forward test

State the mechanism that is supposed to produce an edge and the conditions in which you expect it to disappear. Define the market, timeframe and session before inspecting the results. Use one study ID across the plan, data files and journal.

The rules must allow every decision to be reproduced. Simply saying “enter on a breakout” is not enough: record the price source, decision time and execution method. Keep a copy of the rules or code from before the test. A change made after seeing the result creates a new version; it does not improve the historical assessment of the previous one.

Record the source and coverage of the data and whether it represents bid, ask or mid prices. Match the resolution to the execution rules. Do not treat a future indicator value or a later revision of a release as information available at the time of the signal. A bar covering both SL and TP requires an explicit rule for resolving the outcome or more detailed data.

Separate the sample used to develop rules (in-sample, IS) from the sample used to evaluate them (out-of-sample, OOS). Record the dates and split method. Repeatedly inspecting OOS and adjusting the rules to it erodes its independence. Run the forward test on data arriving after the rules are frozen, documenting all signals and omissions; a demo environment does not fully reproduce execution on a live account.

Define the metrics and decision rule before seeing the result. No single trade count or profit factor threshold confirms an edge for every strategy. Sample length, dependence between observations, rare events and the number of variants tested affect interpretation. Keep the criteria for evaluating the strategy separate from checks that the test itself was performed correctly. Decide in advance how to account for uncertainty and dependence between results, and when to end the study. Do not stop the test at the first favourable result or extend it until the desired threshold is reached; such decisions require a planned method of sequential evaluation or an explicit acknowledgement that further analysis is exploratory.

For consistency with journal F, define a trade’s R-multiple as its net result divided by its positive risk amount planned before entry. The sum of these R-multiples when risk amounts vary is not the result expressed in the plan’s baseline R. The cumulative R curve starts at 0; drawdown is the decline from its previous peak, so an initial loss of −1 R produces a drawdown of 1 R. For equity or balance, specify the starting capital, cash flows and how open positions are treated — do not start a percentage calculation from zero capital. PF with no losses has a zero denominator: mark it as undefined, rather than treating it as confirmation of an edge. Account for partial exits using one documented definition of a trade.

Record results separately for IS, OOS and forward testing. Include counts of signals, trades and exclusions, with reasons. Keep the path of results, not just the total. Net results account for costs exactly once: if spreads and slippage are already reflected in execution prices, do not deduct them again. Mark missing data as missing, not as zero.

Distinguish a fix for an implementation error from a change to the idea. In both cases, keep the previous result and identify the data already seen. Do not overwrite the old protocol version when changing the rules. An export documents the state at a particular time; the file or local save alone is not independent proof of the freeze date.

Compare the result with the criteria in section 5. Record both evidence supporting the hypothesis and facts contradicting it. A conclusion that “more data is needed” requires a concrete plan. This document organises the study; it neither authorises live trading nor guarantees future results.