When Your Trading Plan Stops Working: Diagnose Before Changing It

A plan can disappoint in several ways: hypothetical results turn negative, actual decisions deviate from the rules, costs exceed assumptions, or the routine becomes impractical. These are different problems. Start with the plan version, review period, eligible event set, and the expectation you believe was missed. If you never specified an expectation or uncertainty range, do not invent one after a bad week. A decline can justify reducing exposure or pausing for review without proving that the underlying hypothesis is permanently broken. Separate the immediate decision about risking money from the slower research decision about why results changed.

Reconcile timestamps, instrument identifiers, quotation conventions, missing observations, and account-currency conversions. Verify that each logged trade corresponds to an eligible rule event and that withdrawals or deposits were not counted as trading results. Compare the tested entry timing with the prices available when you actually acted. A setup selected using a completed hourly candle cannot be honestly simulated with a fill that occurred earlier in that hour. If the historical record contains revised or hindsight-dependent inputs, label them. A data or bookkeeping failure can mimic a strategy failure, and changing the trading rules will not repair that underlying measurement problem.

Classify each decision as followed, deviated, or unclear using evidence saved before the outcome. An invented review finds ten decisions: seven followed the frozen rules, two used larger size, and one cannot be classified because the entry note is missing. Review those groups separately without assuming the deviations caused all losses. A compliant losing trade still counts as compliant; an improvised winning trade still counts as a deviation. This prevents profit from rewriting the process grade. Investigate the practical trigger for deviations, such as unavailable monitoring time or ambiguous instructions, so any procedural repair addresses a documented obstacle rather than a moral judgment about discipline.

Suppose a fictional sample of twelve equal-risk-unit trades has five winners of +1.5R and seven losers of −1R. Its gross result is 5 × 1.5 − 7 = +0.5R. If each trade incurs an illustrative all-in cost of 0.08R, total costs are 0.96R and the net result is −0.46R. R here is a fixed planned-loss unit for this simplified exercise; real losses can exceed it. The gross idea and net implementation tell different stories. Inspect whether the assumed spread, commission, slippage, and holding costs match actual conditions. Do not respond to a cost problem merely by adding a new indicator to the entry chart.

Twelve trades provide a narrow and potentially unrepresentative view. Outcomes may cluster because positions share a currency, session, or news event, so the number of rows need not equal the number of independent observations. Statistical methods for proportions make their assumptions explicit; a win-rate percentage alone is not a confidence statement. Also examine the distribution of gains and losses, exposure, and the conditions represented. Do not choose an arbitrary count and declare that crossing it proves an edge. A useful review can conclude that evidence remains insufficient. Continue only within a separately justified risk policy, or use simulation while gathering better observations.

If records and adherence are credible, compare observable conditions under definitions fixed before sorting outcomes. Perhaps execution costs increased, eligible sessions changed, or the sample contains more directional or unsettled periods under your existing classification. Ask whether the earlier research covered those conditions. Avoid inventing a new regime label solely because it isolates the losing trades. Write a mechanism that could connect the condition to the rule, then identify evidence that would contradict it. Several causes can coexist. A plan may be poorly specified, costly to execute, and under-sampled at once, so the review tree should permit unresolved branches instead of forcing a single comforting explanation.

Choose one concrete response tied to the diagnosis: repair the dataset, clarify a rule, change an execution assumption, or stop testing a hypothesis that no longer deserves resources. Save the old version, the reason, and the evidence cutoff. If a substantive rule changes, evaluate it on later untouched observations rather than merging its best results with the old record. Set a review date and a stop condition for the research itself so endless tuning does not masquerade as progress. A careful diagnosis can lead to abandoning the plan. The purpose of the process is an honest decision about the next experiment and exposure, not preserving a strategy at any cost.