Knowing what to buy — and exactly when to act on it.
Two problems sink most active traders: picking the right shares, and catching the right moments to enter and exit positions. Human intuition is poorly suited to either — let alone both simultaneously — when applied consistently across a market-wide universe of daily prices, fundamentals, and news.
Most approaches solve one or the other. Selection models pick what to hold but ignore when to move. Timing models react to price without knowing what's worth holding. The two problems are coupled — getting one right while ignoring the other still loses money.
The goal: automate both decisions in a single pipeline — a model for what to hold and a model for when to act — refreshed every morning before the open on all data through the prior day.
Ranking hundreds of shares for a one-month horizon requires fusing price-action, fundamentals, macro, volatility and options signals — and proving which architecture is actually best, not assumed.
Deciding when to buy, hold or sell is a sequential decision problem with a delayed, noisy reward signal. Supervised learning doesn't apply; the agent must balance short-term gains against drawdown without overfitting to history.
Years of global news needed to become a compact, stable daily signal — without the cost or brittleness of a commercial feed. A commercial source was evaluated and replaced by open data when it proved too limited.
LSTM selection model
A rigorous bake-off against gradient-boosted trees and 1-D convolutional nets established the LSTM as the right architecture for this data. The model ranks the full market by expected one-month return, ingesting price-action, fundamentals, macro, volatility and options signals fused into a structured daily feature set.
Reinforcement learning for timing
The entry/exit agent went through several generations of reinforcement learning, each more stable and risk-aware — evolving from value-based approaches toward policy-based methods with continuous position sizing. Risk-adjusted reward shaping prevents the agent from chasing returns at the expense of consistency.
News as a quantitative signal
Years of global news distilled into daily language-model embeddings, then compressed to a compact signal the models can actually use — adding approximately 12% annual return. Evaluating and replacing a commercial news feed with open source proved the signal was robust and not a data artefact.
Full R&D ownership: model architecture decisions, data engineering, training methodology and end-to-end automation. Every generation was defined against an explicit baseline, cross-validated and benchmarked on risk-adjusted terms — not just raw returns.
In two-year backtests the multi-share portfolio earned 46–47% more profit than buy-and-hold from the same starting capital. Even the broad-market ETF alone beat buy-and-hold by ~39% with the model active. The system now runs a fully automated daily data ingestion, selection and entry/exit decision pipeline — unattended, every morning before the open.