Backtesting crypto has a problem stocks don't

The Asia-session slice is such a specific and real problem. Mine was cutting the funding window.

Easy check: run the same backtest with the boundary shifted six hours. If results move a lot, the boundary is a parameter you never tuned.

Building on Dee's shift test, three things I'd add: 1) run all four shifts, not just one. If only 06:00 looks good you found noise, not an edge. 2) Do NOT then "tune" the boundary to the best one. That's a free parameter with no economic story behind it, and it will overfit faster than almost anything else you can add. 3) If your rule genuinely cares about session structure, stop using a midnight at all and define the boundary off the thing you actually care about, funding stamps, or the open of whichever equity session drives your pair. Then the choice is defensible instead of arbitrary.

Every so often I'm reminded that not having a daily reset to worry about is its own kind of edge. My holds run past every boundary anybody could pick, so the question just dissolves. Sleep is great, you should all try it.

Same trap on the risk side. My daily loss limit resets at a boundary I picked once and never looked at again. Which means a bad run can straddle two "days" and I take double the damage I thought I capped. Added it to the checklist: know where your day starts, and know it's the same boundary for stats and for limits.

Ha, I'm guilty of the local-midnight thing purely because my routine starts with coffee at the same hour every morning and it felt natural. Never occurred to me my numbers were quietly shaped by my sleep schedule. Going to actually write down which boundary I'm using instead of just assuming.

Honest question — does the same thing bite day-of-week features? If my "Monday" starts at UTC midnight but most of the volume I care about happens somewhere else, am I mislabelling a chunk of every week? I've been feeding a day-of-week column into a model without thinking about it at all.