Most retail-level prediction projects train on OHLC candles, which throw away most of what a market was doing. Quantyze starts somewhere else. The limit order book, the live ladder of bids and asks, still holds the structure: supply, demand, and intent. The engine takes order-book snapshots, builds features out of that microstructure (depth imbalance, spread dynamics, level concentration), and feeds them to neural networks trained to predict short-horizon direction.

The pipeline runs in stages: collection, normalization and feature engineering, training, then evaluation against baselines. Evaluation was half the project. Financial data makes it very easy to fool yourself, whether through lookahead leakage, a convenient sample window, or a model that “predicts” by learning autocorrelation that disappears out of sample. I learned to treat a promising backtest as a bug until I’d proven otherwise, and to beat dumb baselines like always-up and momentum before getting excited about anything.

The difficulties were the honest ones for this domain. Signal-to-noise close to zero. Non-stationarity, so a model trained on one regime quietly degrades in the next. Class imbalance in the direction labels. The engineering piled on top of that: high-frequency data volumes, aligning asynchronous book updates into clean training samples, and keeping feature computation fast enough to be worth anything.

Quantyze is more research instrument than product. Its real output was rigor. Splits that respect time, hypothesis first, and a lasting respect for how hard it is to pull real edge out of a market. That habit shows up in everything I’ve built since.