How I validate a trading strategy before it touches real money

Every strategy I have ever tested looked profitable at least once. That is not a compliment to my ideas, it is a warning about backtests. A chart with an equity curve going up and to the right is the easiest artifact to produce in this whole discipline. Producing one that still holds up after honest scrutiny is the actual work.

This is the checklist I run before any strategy, mine or anyone else's, is allowed near real money.

1. Hunt the look-ahead bias first

Look-ahead bias means your simulation used information that would not have existed at decision time. It is the #1 killer, and it is sneaky:

• Signals computed on the forming candle. If your indicator reads the current bar's close before the bar is closed, you are trading with tomorrow's newspaper. Every signal must be computed on closed candles only.
• Execution at the same candle's close. If the signal fires at the close of candle N, the earliest realistic fill is the open of candle N+1. Backtesters that fill you at the signal candle's close are quietly gifting you edge.
• Indicators that repaint. Some indicators recalculate their past values as new data arrives. Their history looks clairvoyant because it literally is.

A useful smell test: if the strategy's win rate drops sharply when you shift execution one candle later, the "edge" was mostly time travel.

2. Charge yourself real costs

Fees, spread, slippage and funding are not rounding errors, on shorter timeframes they are the P&L. A strategy that trades often can go from +40% a year to negative just by charging taker fees and a realistic 0.01,0.05% slippage per side.

My rule: model costs slightly worse than reality. If the edge survives pessimistic costs, reality will be a pleasant surprise. The reverse setup, optimistic costs, unpleasant reality, is how accounts die slowly.

3. Out-of-sample or it didn't happen

Optimizing parameters on all your data and then admiring the result is called curve fitting, and it feels great right until you go live. Split the data:

• Train on one segment, pick parameters there.
• Validate on data the optimizer has never seen.
• If you tweaked anything after peeking at the validation set, congratulations, the validation set is now training data. Get a third segment.

The honest version of this is walk-forward analysis: optimize on a window, test on the next one, roll forward, repeat. The aggregate of all the test windows is the closest thing a backtest gets to telling the truth.

4. Poke the parameters

A real edge is a hill, not a needle. If the strategy is profitable with a 21-period lookback but falls apart at 19 and 23, you have not found an edge, you have found a coincidence with good marketing.

I test every parameter at ±10% of its optimized value. If the neighborhood is profitable, the setting is robust. If only the exact value works, it gets rejected no matter how beautiful the equity curve looks.

5. Paper first, size later

Even after all of the above, the first deployment is paper or minimum size. Not because the math changed, but because live markets add the variables no backtest carries: latency, partial fills, an exchange that goes down, and, most of all, you, watching an open position with real money attached.

The uncomfortable summary

Most strategies do not survive this checklist. That is the point. The filter is supposed to be brutal, because the market's filter is worse. What survives is rarely spectacular, but it is trustworthy, and trust is what lets you keep executing a system on the days it loses.

That, and not any single indicator, is the actual edge.