green rose systems

“growth, forever”
A Software Studio · Established 2026

large-price-model

A transformer on SPY fifteen-minute bars across twenty parallel time series.

01  ·  description

A transformer predicts the next fifteen-minute return of the SPDR S&P 500 ETF (SPY). It reads eight years of fifteen-minute bars (2015 through 2022) across twenty parallel time series. One is SPY itself. The rest carry other stocks, currencies, macro releases, options data, and calendar events. For each bar the model outputs a probability distribution over thirty-two buckets of the next return. The buckets are quantiles of the training-set return distribution, so each holds roughly the same number of training examples. This turns return prediction into bucket classification, the kind of task a language-model architecture (attention over past tokens) is built for.

A normalizer scales each channel using statistics computed only over 2015 through 2022, so the model never sees held-out numbers at training time. A held-out window covering 2024-01 through 2025-06 was set aside before training and may be read at most three times over the project’s lifetime, enforced by a filesystem guard and a commit-message hook.

Three baselines read only the target’s own bucket history — a linear softmax, a three-layer MLP, and a one-layer GRU — each tuned over learning rate and weight decay and stopped at its best validation step. Across five seeds the transformer reaches 3.180; tuned linear 3.256, MLP 3.212, GRU 3.219. Every transformer seed beats every baseline seed. A uniform guess over thirty-two buckets scores 3.466.

The gain lives in the shape of the distribution. The transformer narrows its bet on which bucket the return will land in.

A simulator opens and closes positions based on the predictions and keeps a ledger, reporting Sharpe (average return divided by return volatility, annualized) with an error bar from block-bootstrap resampling. Apache 2.0.

02  ·  repository

github.com/laffeyp/large-price-model

Contact

We’re open to a small number of new engagements. Write with a short note about the work and the timeline.

info@greenrosesystems.com