Pull up your best backtest and look at the lookback. 47 days? 112? Now ask yourself why it isn't 45 or 120.

If the honest answer is "that's where the equity curve looked best," you didn't find a parameter. You found a coincidence and gave it a name.

The number that wins the backtest

Say you test a simple trend rule on BTC. Buy when price is above its N-day average, step aside when it isn't. You try every N from 10 to 250 and plot the results.

You'll get a bumpy line. Some values of N will look great and a few will look terrible. Take the peak and call it your system.

Here's the catch. With 240 candidates and a few years of crypto data, one of them has to come out on top, even if the true edge is the same everywhere. Picking the winner of a lottery doesn't tell you the ticket was special.

Why pros blend instead of pick

As far as I know, the large trend programs rarely bet on one window. They run several at once: short, medium and long. Something like one month, three months and twelve months is a common spread. The academic work on time-series momentum points the same way, with the effect showing up across a wide range of horizons rather than at a single magic one.

That's the tell. If one lookback were truly special, the professionals would have found it and used it alone. They blend because the best window is unknowable in advance, and it changes.

A short window wins in a fast, violent year. A long one wins when the trend grinds for months. You don't get to know which year you're in until it's over.

Overfitting with extra steps

I think this is the part people miss. Optimizing a lookback feels more respectable than optimizing ten indicators, because it's only one number. But it's the same mistake in a smaller coat.

You still looked at the whole history, tried many options and kept the one that fit. The out-of-sample result is usually worse, and the gap is exactly the amount you overfit.

Here's a made-up example to show the shape. A 90-day lookback backtests at 70% a year. The 80 and 100 neighbors do 20% and 25%. That spike isn't an edge. It's noise that happens to sit on 90.

A quick way to feel this: split your data in half. Find the best lookback on the first half, then check how that same number does on the second. Most of the time it slides down the rankings. Sometimes it lands near the bottom. That's the cost of choosing.

The blended version rarely tops either half. It rarely finishes last, either, and that's the property you actually want when real money is on the line.

What to do instead

Look for a plateau, not a peak. Plot results across a wide range of lookbacks. If 60 through 140 all look roughly similar, the idea is robust and the exact number barely matters. If one narrow spike stands out, distrust it.

Then stop choosing. Run three or four windows spread across that plateau and split your risk between them. You give up the fantasy of the best curve. You get a system that doesn't fall apart when the market stops resembling your sample.

Ask yourself this before your next deployment: if the lookback moved 20% either way, would the results change much? If yes, you haven't built a strategy yet. You've built a screenshot.

That's roughly how I approach the rules behind our own systems. We publish the methodology and backtest data at v33systematic.com.