How I test models so I don't fool myself

Zero curve fitting. That's the rule.

Retail trading is full of optimized backtests that look amazing and blow up live. I spent 15 years in data modelling, so I know how easy it is to fool yourself. My rule is simple: if it needs curve fitting to work, it doesn't work.

"No 200% screenshots. No perfect equity curves. Just models that have to survive 10 years of ugly data first."

Built in MQL5, running on a VPS

Everything is custom coded in MQL5 for MT5. No bridges, no copy trading, no signals. It runs on a low-latency VPS so it doesn't spam the broker and timing stays consistent. It's boring, logical code — that's the point. Removes me from the decision.

The 10-year baseline

I don't test on 2 years of nice trending data. Every model has to go through at least 10 years of raw tick data — different regimes, news shocks, black swans, low vol, high vol. If it can't survive that, I bin it.

How I stress it

Walk-Forward: Build on one chunk of data, test blind on the next chunk I never showed it. If it only works on the chunk it saw, it's useless.
Monte Carlo: Randomly shuffle trades and slippage thousands of times to see where it breaks. I want to know the worst case before the market shows me.
Position sizing: I use math (Kelly-based) to keep drawdown capped. I'm not here to double accounts. I'm here to not blow up. I trade own capital, I can't afford to blow up.

That's it. No secret sauce. Just 15 years of learning how not to lie to yourself with data.

→ Next: How I execute with discretion
← Back to Home