README.md

---
license: apache-2.0
viewer: false
tags:
  - finance
  - quantitative-finance
  - factor-model
  - risk-model
  - equities
pretty_name: BUSE1-TH Model Artifacts
---

# BUSE1-TH model artifacts

Released model output for **BUSE1-TH**, an open US trading-horizon equity
factor risk model. Code, methodology and audit suite live in the
[buse-open repository](https://github.com/Binomial-Capital-Management/buse-open).

This repository holds derived model quantities only — no raw market data.
317 monthly point-in-time partitions spanning **2000-01 to 2026-09**
(~1.27 GB, 1,807 parquet files).

> The dataset viewer is disabled: this is a set of six parallel
> monthly-partitioned trees, not a single tabular dataset with
> train/test splits. Load the tree you need directly (see below).

## What changed in v1.1.0

`buse_v1.1.0` is a **survivorship-complete** rebuild of the same model
specification (`risk_model_version` is unchanged, `buse1_th_v1`). The
estimation universe now carries delisted names through their full trading
history, so historical cross-sections are roughly 50% larger than in
v1.0.0 (a typical 2025 exposures partition is ~60,800 rows over ~2,650
tickers, against ~40,800 rows over ~1,777 tickers before). The latest
cross-section covers 2,437 assets as of 2026-09-01. Partition layout,
column schema and hash-manifest contract are unchanged.

The previous release (`buse_v1.0.0`, `imp_v1_0_0`, `retail_v1_0_0`) remains
available in this repository's git history.

## Contents

### `buse_v1.1.0/` — core model

| Tree | Partitions | Coverage | Size |
|---|---|---|---|
| `exposures_by_month/` | 317 | 2000-01 .. 2026-09 | 614 MB |
| `specific_risk_by_month/` | 316 | 2000-06 .. 2026-09 | 496 MB |
| `proxy_exposures_by_month/` | 312 | 2000-09 .. 2026-09 | 112 MB |
| `factor_covariance_by_month/` | 313 | 2000-09 .. 2026-09 | 28 MB |
| `proxy_specific_by_month/` | 312 | 2000-09 .. 2026-09 | 6.3 MB |
| `scale_diagnostics_by_month/` | 313 | 2000-09 .. 2026-09 | 0.3 MB |

Staggered start dates are estimator warm-up, not missing data: specific
risk needs a 5-month EWMA burn-in, proxies and diagnostics 8 months. The
final `2026-09` partition holds the single session available at build
time (2026-09-01); it is a valid point-in-time snapshot, not a full month.

Partitions are Hive-style (`month=YYYY-MM/part.parquet`) and hold **daily**
point-in-time rows x 122 columns, plus the `month` partition key when read
through `hf://`. `scale_diagnostics_by_month/` is one JSON file per month.

Also at the tree root: `factor_returns.parquet` (full daily history),
`exposures_latest.parquet`, `specific_risk_latest.parquet`,
`factor_covariance.parquet`, `factor_universe.json`, `metadata.json`,
`hash_manifest.json`, `replicated_carrier_exposures.parquet`,
`replicated_carrier_report.json`, `audit.md`.

`audit.md` is the build-time audit for this generation as run. It records
one FAIL (`E.proxies`: the QQQ anchor's projected beta of 1.37 sits outside
the audit band) that was reviewed and accepted before release; the other
21 checks pass or flag informationally. The v1.0.0 calibration matrix
(`calibration_report.json`) has not yet been re-run on this generation.

### `imp_v1_1_0/` and `retail_v1_1_0/` — specific-risk overlays

Base-agnostic monthly tables of `[date, ticker, variance_multiplier]` with
hard bounds `[0.25, 4.0]`. Each tree carries `metadata.json` and a hash
manifest.

| Overlay | Partitions | Coverage |
|---|---|---|
| `imp_v1_1_0/` | 128 | 2016-01 .. 2026-08 |
| `retail_v1_1_0/` | 104 | 2018-01 .. 2026-08 |

- **IMP** — per-ticker multiplier from the cross-sectionally centered
  implied/realized volatility ratio. Zero fitted parameters.
- **RETAIL** — multiplier from abnormal attention not yet confirmed by
  volume (Wikipedia attention z-score minus volume z-score), with one
  calibration constant frozen before out-of-sample evaluation.

Both overlays are rebuilt with the same frozen rules as v1.0.0, extended
by one month. The out-of-sample gate reports published with v1.0.0
(`evaluation_report.json`) were not re-run for this rebuild and are not
included here. Overlay inputs are not survivorship-complete: names that
delisted before the rebuild ride the base model (the documented
missing-symbol policy).

## Usage

Download the full tree, or just the part you need:

```bash
pip install huggingface_hub[cli]

# everything (~1.27 GB)
hf download BinomialTechnologies/buse-artifacts \
    --repo-type dataset --local-dir artifacts

# one overlay only
hf download BinomialTechnologies/buse-artifacts --repo-type dataset \
    --local-dir artifacts --include 'imp_v1_1_0/*'
```

Read a single partition without downloading the rest:

```python
import pandas as pd

url = ("hf://datasets/BinomialTechnologies/buse-artifacts/"
       "buse_v1.1.0/exposures_by_month/month=2025-10/part.parquet")
exposures = pd.read_parquet(url)
```

Apply an overlay to any risk model's specific variances:

```python
from buse1.overlays.loader import apply, load_overlay

mult, provenance = load_overlay("artifacts/imp_v1_1_0", decision_date)
conditioned = apply(base_specific_variance_by_ticker, mult)
```

## Integrity

`SHA256SUMS` covers all 2,132 artifact files:

```bash
cd artifacts && sha256sum -c SHA256SUMS
```

## License

Apache-2.0. These are derived model quantities; see
[docs/limitations.md](https://github.com/Binomial-Capital-Management/buse-open/blob/main/docs/limitations.md)
for data provenance and reproducibility boundaries.