Measure your live-vs-backtest slippage before you scale anything

Seconding the distribution point. Mean slippage is the least informative statistic in the whole exercise.

This bit me on thin names specifically. Same strategy, liquid names fine, small caps a totally different story.

Adding to the checklist: log slippage per fill from day one, not retroactively. And tag each fill with size and time of day. If your tail is all concentrated in the first five minutes of the session, that's a scheduling fix, not a strategy fix. Also please don't scale on a 50-fill sample. That's a hint, not a number.

Learned this one the expensive way back when a fill was something a person on a phone did for you. The strategy on paper was lovely. The strategy after the guy on the other end knew what I wanted was a different animal entirely.

Genuine question on the tail — how are you separating slippage that's a cost of being in the market from slippage that's your own order moving the book? Same fifty fills, but I'd expect the second kind to scale with your size and the first not to. Would halving order size for a stretch and re-measuring be a fair test of that?

Amusing that the thing everyone tunes for months is the entry signal and the thing nobody measures is the price they actually got. Backtest fills are a work of fiction and we all just quietly agree to believe it.

Crypto version of this is fun because the spread changes shape depending on what hour you're awake. My overnight fills and my daytime fills basically belong to two different strategies, so I keep them in separate logs now.

Ok so honest question, when I train on backtest prices, the model is learning against fills that never existed, right? Should I be feeding it my real slippage distribution somehow, or is that overcomplicating it before I even have enough live fills to be meaningful? Still pretty confused about where that belongs.

One thing worth adding on the sizing side: slippage doesn't just shave your returns, it shaves the input you size off. If your per-trade edge drops from say 20bps to 12bps after real fills, your Kelly-ish fraction drops by the same proportion. Edge sits in the numerator, variance in the denominator, and the variance barely changes. So an 8bp execution cost can cut your justified position size by 40%, which is a much bigger deal than the 8bp sounds. @MLcurious, before you feed anything into a model, just subtract a pessimistic slippage constant from every backtest trade and see if the strategy survives at all — if it doesn't, the distribution work is moot.