Most retail-level prediction projects train on OHLC candles, which are heavily lossy summaries of what a market is doing. Quantyze starts from a different premise: the limit order book — the live ladder of bids and asks — contains structural information about supply, demand, and intent that candles throw away. The engine ingests order-book snapshots, engineers features from that microstructure (depth imbalance, spread dynamics, level concentration, [order-flow deltas]), and feeds them into neural networks trained to predict short-horizon price direction.

The pipeline runs in stages: data collection from [exchange API], normalization and feature engineering, model training, and evaluation against baselines. Getting the evaluation right was half the project — with financial data it is incredibly easy to fool yourself. Lookahead leakage, survivorship in your sample window, a model that "predicts" by learning autocorrelation that vanishes out of sample. I learned to treat any promising backtest as a bug until proven otherwise, and to benchmark against dumb baselines (always-up, momentum) before celebrating anything.

The struggles were the honest ones of the domain: signal-to-noise near zero, non-stationarity (a model trained on one regime degrading in the next), and class imbalance in directional labels. The technical struggles compounded them — handling high-frequency data volumes, aligning asynchronous book updates into clean training samples, and keeping feature computation fast enough to matter.

Quantyze is less a product than a research instrument. Its real output was forcing rigor: proper train/validation/test splits across time, hypothesis-first experimentation, and respect for how hard it is to extract real edge from markets. That rigor now shows up in everything else I build.