Spent two years trading crypto manually — chart-watching, FOMO entries, revenge trades after losses. My win rate was fine, my behavior wasn't. Emotion was costing more than bad setups ever did.

So I trained XGBoost models on OHLCV data, funding rates, and order flow imbalance across major pairs, and let the model — not my gut — decide entries, targets, and stop-losses. A few things I learned building this:

• Feature engineering matters more than the algorithm. Raw price data alone barely beats a coin flip. Volatility regime features and funding rate deltas were the biggest signal boosts.

• XGBoost handles regime shifts better than I expected, but it still needs retraining on a rolling window — crypto's correlation structure breaks every few months.

• The real edge isn't prediction accuracy, it's consistency. A model with a 55% hit rate that never panics beats a trader with a 65% hit rate who blows up once.

If you're a retail trader tired of emotional entries and want to see what disciplined, data-driven signals look like day to day, I opened it up as AlphaBoost — happy to answer questions on the modeling approach in the comments.