A strategy that lasts
Most strategies that look good in a backtest stop working live. The usual reason is overfitting: the rules and parameters were shaped around one stretch of history, and fit it better than they fit the market.
Start from an idea you can say
"Stocks that gap down on no news tend to recover by the close" is an idea. "RSI 13, stop 1.7%, only on Tuesdays" is a result of searching. If you can't say why a rule should make money, a good backtest of it is probably luck.
Treat one backtest as a first look
A single backtest is exploratory. Its numbers are always optimistic, because you chose the strategy while looking at that same past. Plan the walk-forward before the first one, and keep your backtests on the in-sample months. See Backtests.
Validate on data the strategy has not seen
Parameters chosen on the data they are scored on are overfitted by construction. A walk-forward scores them on the next window instead. Keep it honest:
- Decide the plan before tuning.
- Use at least 10 folds; 12 months in-sample and 3 out-of-sample is a sound default.
- Never go back to a closed fold, and never peek at out-of-sample dates.
- If you ran several plans, the best one is not the validated one.
Ask for enough trades
A result on 20 trades says little. The Edge Report asks for at least 100 out-of-sample trades before it calls a strategy ready.
Watch the parameters
If the best parameters jump around from fold to fold, the strategy is chasing noise, even when the total looks good. Fewer parameters make this less likely.
Count the costs
Unless the strategy says otherwise, a backtest assumes no slippage and no stock commission. Ask the AI to add costs that match your broker, and see whether the result survives.
Trust the numbers made here
A backtest result found on the web was made on someone else's data and rules. Run the idea here before believing it.
Go live small, on paper first
Then compare the live with its shadow backtest. See From backtest to live.