When the Backtest Lies: Twenty-One Data and Engine Errors from a Gold Trading Research Pipeline

This note documents twenty-one errors found during 2026 in a research pipeline for intraday gold strategies: an event-driven Python backtest engine fed by a MetaTrader 5 broker, Dukascopy and CME futures data. None of them crashed anything or produced an obvious warning. Each produced plausible numbers, and several reversed a conclusion: a strategy with a 79% win rate and a profit factor of 12 had no edge at all; another went from profit to loss when one data feed was replaced by another; a month that seemed to break a strategy was an hour of misaligned clocks; a profitable trade around US data releases was rejected because its window was set in UTC; and a plan to move to futures rested on a spread estimate three years out of date. For each error the note gives what happened, how it was found and how much it moved the result, and it closes with the checks that caught them. Errors that appear only when a backtest is compared with live trading are covered in a companion note. AI disclosure. The author used Claude (Anthropic), an AI assistant, to write the data-analysis code, todraft and edit the text, to assist with wording and structure, and as a consultative tool for discussingmethodological and technical questions. The research questions and the decisions on what to test and whatto accept are the author’s. The author reviewed, verified, and edited all AI-assisted content and takes fullresponsibility for the content of the paper, including its analysis, results, and conclusions.