Markets
- Question
- Does a self-supervised model of raw trade and quote data learn structure that hand-built volatility features miss?
- Status
- Research programme, stopped at a documented verdict. Mixed result by design: it beats the best named baseline on some targets and loses on others.
The programme records hypotheses and decision rules before runs. Baselines are named in advance, results are reported on two separate eras, and a result only counts if it holds in both.
- A causal transformer of about 20 million parameters (grouped-query attention, rotary embeddings, SwiGLU) on per-symbol dollar-volume "states", with a next-state objective, a quantile head and cross-channel coupling, read out frozen through one linear layer.
- An assembled corpus of hundreds of trading sessions across 2018 to 2025 and over a thousand tickers. Sub-penny placement, odd-lot, venue and intermarket-sweep flags are first-class input channels, and a per-symbol dollar-volume clock is the model's native time axis.
- Named specialist baselines (HAR-RV and its realised-quarticity and volume-clock variants, volume-profile and range models), strengthened step by step with the cost recorded. Foundation-model adapters (Chronos-2, TimesFM, Toto) are implemented for further comparison.
- Controls: a symbol-identity leakage check, and a single-channel ceiling that asks whether one raw input matches the encoder.
- A companion study on whether market histories are better grouped by what followed them than by what they look like. It closed as a negative result.
What I found
- Two of five targets beat the best named baseline in both eras. Three did not, and the spread and depth target lost clearly in both.
- Unanimity across six splits within one era is not replication, so every claim needs both eras.
- A window of about six minutes of states matched one of nearly an hour, which says something about the problem and not only the model.
Earlier: a QuantConnect strategy
A long-only equity strategy that ranks S&P 500 stocks by a crash-risk measure (down-to-up volatility, after Chen, Hong and Stein, 2001), tracked weekly over several time frames together with its speed and acceleration. Entries come from a signal ladder, and a scored daily check handles exits. Backtested from 2019 on a cash account.