A 26,213-parameter numeric transformer that estimates the next official
short-interest settlement value from the completed FINRA daily short-volume
window. Each token is one trading day of numeric data, not text.
The model outputs P10, P50, and P90 estimates for the log change from the latest
reported short interest, plus an uncalibrated cross-sectional ranking score. It
predicts the next settlement endpoint—not an authoritative daily series—and
does not equate short volume with open short positions.
ONNX inputs
Input
Dtype
Shape
daily_features
float32
[batch, 32, 21]
aggregate_features
float32
[batch, 28]
valid_days
bool
[batch, 32]
Normalization is embedded in model.onnx; do not normalize the engineered
inputs again. Exact feature order is recorded in model-config.json and
manifest.json. The four outputs are
median_log_change, p10_log_change, p90_log_change, and rank_score, each
with shape [batch]. Convert a log-change output delta back to shares with:
Clamp operational endpoint estimates at zero after the inverse transform.
Training and results
Training used 660,246 consecutive settlement windows from January 2020 through
July 15, 2026, combining official short-interest endpoints, FINRA daily volume,
and filing-date-aware SEC shares outstanding. A two-layer, width-32, four-head
transformer was selected across seeds 17, 29, and 43 with AdamW (batch 512,
learning rate 0.0005, weight decay 0.0001). Development used data through 2023
for training, 2024 for selection, and 2025 through July 15, 2026 for the
retrospective test. The released checkpoint was fit from scratch for seven
epochs on all data with seed 17; the metrics below belong to the frozen
three-seed evaluation, not that final checkpoint.
Method
Pooled Spearman
Mean per-date Spearman
Ratio-change baseline
0.2751
0.2755
Raw-feature ridge
0.3428
0.3452
Numeric-transformer ensemble
0.4143
0.4145
Limitations
Evaluation uses retrospective final-file FINRA data rather than strict
point-in-time vintages, and the test period is not a pristine preregistered
holdout. There are no authoritative daily short-interest labels. Training data
and source are private; the published weights and metadata are Apache-2.0. This
research model is provided without warranty and is not investment advice.
