Learn
Backtest the way a prop firm scores you
A backtest that only measures profit is lying to you about a prop account. The firm does not score you on net profit. It scores you on whether you ever touched the drawdown, whether you cleared the consistency rule, and whether you were flat by the cutoff. So you backtest against those, not just the equity curve. This is how to do that, and the traps that make a bad strategy look good.
I backtest my own bot against the firm's constraints before it trades a cent, because the gap between a clean backtest and a surviving funded account is where most bots quietly die.
Cap the equity curve at the firm's drawdown
The single most important change: run the backtest with the firm's drawdown enforced as a hard line, the same way the firm's risk engine will. If the account is a 50K with a 2,000 dollar trailing max loss, the moment your modeled balance drops 2,000 from its trailing peak, that account is dead in the test. Not paused, dead. Then you start a fresh one and count it as a reset you paid for.
Most off-the-shelf backtests do not do this. They show you a smooth net-profit line that quietly went 2,500 dollars underwater in March and recovered. On a real prop account that line ended in March. A strategy that is net profitable but breaches the drawdown once a quarter is a losing prop strategy, and only a drawdown-capped backtest shows it.
Model the payout gates and the flat-by time
Add the rules that decide whether you actually get paid. The consistency rule, often that no single day can be more than 30 to 50 percent of total profit, changes which strategies pass, because one huge session can lock your payout. The minimum trading days and any profit buffer above the starting balance change how long until the first withdrawal. And the flat-by time means any position the bot holds past the cutoff is closed for you, so trades that straddle it have to be modeled as forced exits.
Backtest these as pass-or-fail conditions, not footnotes. A strategy that hits the profit target in two big days fails a consistency rule that a slower, steadier strategy clears easily, even if the slower one makes less.
The traps that fake a good result
Lookahead bias is the worst: the backtest uses information the bot would not have had in the moment, like the day's high to time an entry. It produces beautiful results that vanish live. Then no slippage and no commissions, which on a futures scalper can be the whole edge. Then sizing that ignores the drawdown, which inflates returns by risking more than a prop account can survive. An AI model will write all three by default unless you tell it not to.
Overfitting is the quiet one. Tune enough parameters and any strategy fits the past perfectly and predicts nothing. Test on data the strategy was not tuned on, and be suspicious of any result that needs an exact setting to work. If moving the stop two ticks breaks it, it was never real.
Using AI to build the harness
Claude or ChatGPT will write a backtest harness quickly, and that is a good use of them, as long as you specify the firm's rules as constraints up front: cap the drawdown, charge slippage and commission, enforce the flat-by time, count resets. Ask for the constraints explicitly or the model defaults to the optimistic version. Then check the numbers by hand on a few trades before you trust the whole run.
The full firm-aware harness I run, the one that scores a strategy against drawdown, consistency, payout timing, and reset cost in one pass, is the part I keep for the paid build doc. The method above is the free version, and it is enough to stop a bad strategy before it costs you an account.
FAQ
How do you backtest a strategy for a prop firm specifically?
Run the backtest with the firm's rules enforced, not just profit. Cap the equity curve at the firm's drawdown so any breach ends that account, model the consistency rule and payout gates as pass-or-fail conditions, force exits at the flat-by time, and count the resets you would have bought. A strategy that is net profitable but breaches the drawdown still fails a prop account.
Can ChatGPT or Claude backtest a trading strategy?
They can write a backtest harness quickly from a clear set of rules, which is a good use of them. The catch is that models default to optimistic backtests with lookahead bias, no slippage, and drawdown-blind sizing, so you have to specify the firm's constraints explicitly and check the numbers by hand before trusting the result.
What is lookahead bias in a backtest?
When the test uses information the bot would not have had at the time, such as the day's high or close to time an entry that happened earlier. It produces excellent backtest results that disappear in live trading, because the live bot cannot see the future. It is the most common reason a great backtest fails live.
Why does my backtest look profitable but the account still fails?
Usually because the backtest measured net profit while ignoring the drawdown, the consistency rule, the flat-by time, slippage, or reset costs. A smooth profit line can hide a moment where the account went past the trailing max loss, which on a real prop firm ends the account regardless of the later recovery.