Research

What we tested, and what failed

Working notes from building LATENT. Mostly about things that did not work, because those are the results nobody publishes.

Draft — not published

Every stop loss I tested made things worse

Four stop-loss variants tested against nine years of BTC/USDT. All four lowered risk-adjusted return.

Read → Draft — not published

What a deflated Sharpe ratio actually corrects for

Trying 730 configurations and reporting the best is how backtests lie. What deflation does about it.

Read → Draft — not published

The bug that made my model wrong on every intraday chart

Daily-fitted constants applied unchanged to intraday bars left the Kalman gain flat and GARCH persistence 6x to 24x too fast.

Read → Draft — not published

The drawdown number lies: 14.5% max, 713 days underwater

Maximum drawdown is the figure everyone quotes and it hides the thing that actually makes people quit.

Read → Draft — not published

Trading more often cost $87,000 and earned nothing

The same signal on four timeframes: five times the trades, four times the fees, the same return.

Read →