How to Backtest Futures Trading Strategies Correctly: Avoiding the Most Common Mistakes

A backtest is only as honest as its assumptions. The usual reason a strategy that looked great on historical data fails live isn't bad luck. It is that the test quietly used information the trader wouldn't have had, tuned itself to noise, or ignored the cost of trading. Here are the errors to check for before you believe any result, including your own.

Lookahead bias

Lookahead means the test uses information that wasn't available when the trade would have been placed. The subtle version: a signal depends on a bar's close, but the test enters at that bar's open or at a price inside the bar. In live trading you can't know the close until the bar closes, so the earliest realistic entry is on the next bar. Check every condition and indicator for what it would have known at the instant of the order, and force entries to happen after the information exists. A strategy that gets noticeably worse when you delay entry by one bar is a warning sign.

Overfitting

Overfitting happens when parameters are tuned until they match one stretch of history, including its random noise. A strategy with many optimized parameters and a smooth equity curve should make you suspicious, because more parameters means more ways to fit noise. The standard check is out-of-sample testing. Optimize on one part of the data. Then run the chosen parameters once on a part you never touched. If you look at the out-of-sample result and go back to adjust, it is no longer out-of-sample. A large gap between the two results suggests the in-sample number was mostly fit. A small gap doesn't prove an edge, but a large one is a clear reason to stop.

Data and market structure errors

Futures contracts expire, so long histories are built by joining contracts. How the join is done matters. Back-adjusting changes historical prices, which affects any rule that uses price level or distance. Unadjusted joins leave a gap at each roll that can create fake trades. Roll dates and the contract you treat as front month vary by product and data vendor, so don't assume yours match another source. Also check session definitions: the hours your data covers, the daily break, and holiday short sessions. Tick size, typical volume and market behavior have changed over the years, so a fast strategy tested over a long history assumes conditions that weren't constant.

Cost modeling, with a worked example

A test with zero commission and fills at the signal price overstates every strategy, and the more often you trade, the more it overstates. Use your broker's actual commission and exchange fees, because those change. Model slippage as a deliberate assumption, not an afterthought. A market order may fill worse than the signal price, and a limit order may not fill at all.

Hypothetical example, for arithmetic only. An ES strategy averages $15 gross profit per trade. If all-in round-trip costs are $10, net is $5 per trade. Add one tick of slippage ($12.50 on ES) and net becomes negative $7.50 per trade. At 20 trades a day, one tick of slippage per trade is 20 ticks, or $250 per contract per day, which is enough to turn a profitable test into a losing one. Run your strategy at zero slippage, one tick and two ticks. If the result disappears at one tick, you have learned something important before spending any money.

One more limit: a backtest or simulated fill can't reproduce your position in the exchange queue or the real size available at your price. A simulated fill isn't validation of live execution.

Walk-forward testing

Walk-forward analysis repeats the optimize-then-test process across rolling windows: optimize on the first block, test on the next, slide forward, repeat. You end up with several out-of-sample results instead of one. If performance is acceptable across most windows and the optimal parameters stay reasonably stable, that is better evidence than a single split. If the best parameters jump around from window to window, or only a few windows carry all the profit, the edge is weak or tied to one regime.

NinjaTrader 8

The NinjaTrader 8 Strategy Analyzer, opened from the Control Center, supports backtests with commission and slippage settings and includes a walk-forward optimization option. Check the official NinjaTrader 8 help guide for the current steps and settings. Its fills are simulated, so the cost caveats above still apply. For how YMI describes its bot logic, refer to the Marty bot specs instead of assuming results from any backtest.

What counts as enough

No single threshold proves a strategy. A short test window, 30 trades, or even 50 simulated trades isn't a proven edge, and simulated trading isn't live proof. A more useful checklist: the data covers more than one kind of market, entries respect lookahead, costs and slippage are realistic, out-of-sample or walk-forward results hold up, and you can explain why the pattern should exist based on market structure. If any of these fail, don't size real capital to the number. To sanity-check expectancy from win rate and reward-to-risk, the edge calculator is a quick tool, though it only works with inputs you have already validated.