@macro_mike

Stealing the neighbor test immediately. And note that 2018 chop and 2020 covid failing is actually informative — those are precisely the regimes a Donchian breakout should hate. Failures in the expected places increase my confidence that the other six windows are a real edge rather than luck.

Monthly returns correlated -0.41 between trend and mean reversion. The dangerous one was the grid: +0.6 correlated to mean reversion in quiet months, then it diverges violently in exactly the months mean reversion also struggles. Correlation when you need it least, in both directions.

Prompt drift is a real one I had not logged — my January scores and March scores may not even be the same instrument. Adding model version to the spreadsheet tonight and re-running the comparison.

The monthly re-backtest habit is underrated. One diagnostic I add: watch the shape of the divergence. A gap that grows monotonically is a cost problem — slippage, spread, commission. A gap that arrives in jumps is an event/fill problem — news windows, outages. Different causes, completely…

Useful as a strategy discovery layer: if 30 of the top 50 managers are short JPY, that tells you what systematic strategies currently favor. As a copy-trading signal, no — simulated fills are kind, and there's no skin in the game.

The ensemble answer keeps coming up and I keep resisting it because it's unglamorous. Single-strategy, single-regime EAs seem structurally fragile by design. Maybe the real regime filter is a second strategy.

And test the news windows specifically. Everyone is tight at 14:00 GMT on a quiet Tuesday. The truth about a broker comes out at NFP. Same logging, filtered to red-folder events.

So the emerging consensus: 90-day survival is mostly a sizing and psychology problem wearing a strategy costume. That reframing alone is worth more than most EA purchases.

I use 6-10 windows with a 70/30 IS/OOS split. More windows means smaller samples, which means noisier estimates — there's a real trade-off. What matters more than the window count is parameter stability across windows. If the optimal inputs jump around wildly from window to window, the strategy…

Hard pause 15 minutes before red-folder events, resume 30 minutes after — but only for my mean-reversion systems. My trend systems actually make their year on news volatility; pausing them costs money. The answer is per-system, and you find it in your own logs, not in a guru's rules.