How to Backtest a Gold Trading Strategy Without Fooling Yourself

Write an executable specification first

A strategy is more than a picture of a good entry. Specify the instrument, feed, timeframe, signal timestamp, entry order, size, exit, session restrictions and handling of overlapping positions. If two reasonable programmers would produce different trades from the description, the rule is incomplete. Version the specification before testing so later improvements cannot silently rewrite the original experiment.

Audit the candles before calculating performance

Check timestamp order, duplicates, conflicting rows, timezone and OHLC consistency. Distinguish weekends and scheduled closures from unexplained missing bars. Keep a gap report and preserve the original provider data. Futures continuous series need a roll and adjustment policy; broker spot derivatives need quotation-side and dealing-hour information. Never splice providers silently just to remove a visible hole.

Respect the moment information became available

A completed bar’s high, low and close are unavailable at its opening. A swing requiring future confirmation cannot be traded at the earlier turning point. Revised economic data may not match the first release. Indicators can recalculate or repaint. Store signal-generation time and use only inputs available at that time. A beautiful historical result built on future information is not an executable result.

Model ambiguous bars explicitly

If a bar reaches both the target and stop, OHLC alone does not reveal which occurred first. Use finer data when available or report a conservative rule and alternative bounds. Do not always award the target first. Likewise, a price touching a limit is insufficient proof that the entire intended size filled. State the fill assumption and test how sensitive the outcome is to it.

Separate development from evaluation

Use an earlier period to develop the rule, freeze parameters, and evaluate on a later untouched period. Repeatedly revisiting the holdout turns it into development data. Keep a record of all attempted variants, including failures. Walk-forward evaluation can help organize sequential testing, but it does not eliminate selection bias or guarantee that future market conditions will resemble the sample.

Report costs and more than the win rate

Include spread, commissions, slippage, financing and roll assumptions where relevant, without double-counting embedded costs. Report trade count, average net win, average net loss, net expectancy, exposure and drawdown. For example, 40% wins averaging 2R and 60% losses averaging 1R imply 0.2R gross expectancy before costs. It is invented arithmetic, not evidence of a profitable strategy.

Know what the site can actually calculate

The current Research Lab supports browser-local candle analysis and event-window studies. It does not by itself provide a complete execution-aware NSI or CY66 strategy backtester. Indicator signals need archived timestamps, versioned rules and defined evaluation logic. Hypothetical results should remain clearly separated from actual trading records; CFTC guidance highlights their limitations. Continue with a trading journal and forward observation before making performance claims.

Follow a connected reading path