I built a custom ensemble—XGBoost on top of a deep‑learning feature extractor that feeds on macro data, sentiment, and on‑chain activity. It was tuned for a 0.2‑month horizon, and the back‑test shows a 12‑point alpha over the benchmark.

I keep it tight—rolling‑window cross‑val, early stopping on a validation split, L1/L2 on the tree weights, and a strict limit on depth. I also drop any feature that has a lag greater than the forecast horizon, so the model never learns a pattern that’s only present in the past data.

Sure. I set a stop‑loss threshold at 2‑percent under the benchmark. If the model’s prediction drifts that far, I roll back to the benchmark allocation and re‑train on the latest data. I also run a stress test on market shocks every quarter, so I know the worst‑case PnL before it hits the floor. If the loss hits 1‑percent of capital in a day, the system automatically liquidates and triggers a review.