Gator Quant Hacks 2026: trading on what the Fed says

A systematic trading study for the Gator Quant Hacks 2026 Systematic Trading track. It asks whether Federal Reserve communication (speeches, statements, minutes and press conferences) predicts next-day moves in long Treasuries and the US dollar. A walk-forward language model scores every Fed document for hawkish or dovish stance. The main strategy, v2 T4xE1, turns those scores into a daily, volatility-targeted position in TLT and UUP. Every test was pre-registered, with a hashed and timestamped plan, before its results were computed.

Bottom line. The strategy failed its own pre-registered decision rule. Its in-sample Sharpe at twice the modelled costs was 0.387, against a required 0.5, and that verdict stands. The two-year out-of-sample window was opened once, for reporting only. There it earned a modest, statistically insignificant result. Adding it to an existing three-sleeve portfolio raised that portfolio's Sharpe, but a paired bootstrap cannot tell the gain from zero. The press-conference voice and face studies found nothing. The full write-up is docs/report/REPORT.md.

Results at a glance

Main strategy: v2 T4xE1 with both D-4 implementation fixes. Returns are daily excess returns over the T-bill, net of modelled costs. In-sample is 2016-01-04..2024-10-02 and out-of-sample is 2024-10-03..2026-10-02.

Every number above was reproduced on HiPerGator. Sources:

• metrics: docs/results/sharpe-summary.md and results_2/metrics/v2_metrics.csv;
• bootstrap: results_2/attribution/bootstrap.csv;
• the original, unfixed results: backtests/results/v2_oos/.

Repository structure

.
├── README.md                 this page: overview, structure, setup, how to run, data
├── docs/                     ALL documentation (index: docs/README.md)
│   ├── report/               the full team report (REPORT.md) and its earlier draft
│   ├── submission/           competition texts: Devpost description, submission status, reproduction commands
│   ├── guides/               how-tos: data, running the backtests, reproduction, HiPerGator, stance model, v2, replay
│   ├── results/              results write-ups: strategy results, Sharpe summary, H2-H4 local run, reviewer-gap coverage
│   └── research/             planning notes, reviews, literature and data-feasibility notes
├── preregistration/          hypotheses, amendments, deviations D-1..D-4 and PREREG_LOG.md (frozen; code checks their hashes)
├── backtests/                the maintained backtest code and every committed result
│   ├── run_all_backtests.sh  one command for every stage: chrono, v2, presser, h234
│   ├── v2/                   the v2 strategy (run_v2*.py, v2lib.py, sim_fixed.py, run_v2_d4.py), its scores, public inputs/
│   ├── presser/              press-conference study: events, Q&A text, ADDENDUM.md, backtest code, H2-H4 reference run
│   ├── wire/                 stance-score wiring and the combined summary
│   ├── tools/                compare_results.py, the reproduction check
│   └── results/              v2/ (in-sample), v2_oos/ (D-3), v2_d4/ (D-4), presser_h1/, presser_strategy/, capacity/, summary.md
├── results_2/                reviewer-gap analyses built from committed outputs only (no new backtest)
│                             metrics/, attribution/, turnover_capacity/, provenance/, replay/
├── nlp/                      walk-forward chrono-BERT stance model: label dating, training, scoring
├── hpg/                      HiPerGator Slurm jobs; fedpress_pkg/ is the voice, face and speech feature package
├── data/                     public data: Fed text corpus, press-conference transcripts and captions, local feature tables
├── strategies/01/            the original strategy 01 hypothesis and its independent review (historical)
├── massive-8k/               the separate Massive 8-K options challenge (notebook and write-up); not part of the Fed study
├── backtest_snapshot/        frozen code, inputs and outputs behind the 2026-10-03 results (provenance record)
└── archive/                  collected session outputs and earlier work (historical, not maintained)

Code and documentation are kept apart. Every guide and write-up is in docs/. Markdown files outside docs/ are one of four kinds:

• records the code reads or hash-checks: preregistration/, backtests/presser/ADDENDUM.md;
• files written by the scripts next to their results: the READMEs, metrics.md and tables.md under backtests/results/ and results_2/;
• frozen history: backtest_snapshot/, archive/, strategies/01/;
• the feature package's own manual in hpg/fedpress_pkg/;
• the separate Massive 8-K challenge in massive-8k/, whose README and write-up sit beside its notebook.

backtests/README.md, backtests/v2/README.md and nlp/README.md are one-line pointers, kept because the scripts' messages mention those paths.

Setup

Python 3.12 and bash (Git Bash on Windows works):

git clone -c core.autocrlf=false https://github.com/savioxavier/gator-quant-hacks.git

cd gator-quant-hacks && python -m venv .venv && source .venv/bin/activate

On Windows Git Bash, activate with source .venv/Scripts/activate instead.

pip install -r hpg/backtest_requirements.txt

hpg/backtest_requirements.txt pins the package versions that produced the committed results: numpy, pandas, pyarrow, scipy, statsmodels and matplotlib. Clone with core.autocrlf=false. The pre-registration hash checks fail if git rewrites line endings.

How to run

There are three levels. Each one needs more inputs than the one before.

1. Replay the published results (public, no licensed data, about a minute)

python results_2/replay/replay.py

This recomputes the headline tables and figures from the committed daily return series and trade rows, then compares them with the committed tables. It covers:

• the D-3 and D-4 v2 metrics;
• the in-sample selection;
• the press-conference strategy tables;
• the G3 headline;
• the H2-H4 Holm family;
• the results_2 metrics table.

It prints PASS or FAIL per item. The exit code is 0 when everything passes, 1 on a mismatch and 2 on a missing input. Output goes to results_2/replay/replay_out/, which is git-ignored. This checks the arithmetic from the published series to the published tables. It does not re-run signal construction or execution. Details: docs/guides/saved-result-replay.md.

The reviewer-gap analyses rebuild from committed files the same way. Each refuses to run on a missing input:

python results_2/metrics/compute_metrics.py && python results_2/metrics/test_drawdown.py

python results_2/attribution/attribution.py && python results_2/turnover_capacity/capacity.py && python results_2/provenance/provenance_check.py

2. Full backtest reproduction (needs licensed market data)

bash backtests/run_all_backtests.sh

The runner has four stages. bash backtests/run_all_backtests.sh v2 (or presser, h234) runs one alone:

What each variable points at:

• V2_WORK_ROOT: $PWD/backtests/v2/inputs/work in this repository.
• GQH_DATA_DIR:
  • the public FRED and T-bill files from backtests/v2/inputs/data/;
  • the private ETF and futures daily files (see Data below).
• GQH_REPO: the backtest engine (gqh-flow-clock at commit 1587f25).
• GQH_MARKET_DIR: the three Databento intraday files.

Re-run outputs go to backtests/rerun/, which is git-ignored and holds price levels, so never commit it. The exit status is non-zero when a stage fails, is refused or does not reproduce. A skipped stage is reported as skipped. Count a run as a reproduction only when every stage you need reports that it reproduces. Full guide: docs/guides/running-backtests.md. What is reproducible from what: docs/guides/reproduction.md.

3. On HiPerGator (UF)

Log in yourself (GatorLink and Duo), clone the repository on /blue, and submit from the repo root:

Guides:

• jobs and the feature package: docs/guides/hipergator.md and hpg/fedpress_pkg/README.md;
• the stance model: docs/guides/stance-model.md.

Data

The repository is public. Licensed prices and quotes are never committed, and that includes results that would reveal price levels. What is published is returns, tick moves, spreads in ticks, P&L and statistics. docs/guides/data.md covers:

• the exact Databento and Yahoo requests;
• the SHA-256 of every private file used, so a rebuilt copy can be checked;
• how the public data were collected;
• what each committed data folder holds.

Documentation

Ground rules this repository keeps

• Pre-registration first. Plans are hashed and committed before results. Deviations are written down before what they authorise is computed. A failed decision rule stays failed.
• The out-of-sample window was opened once (deviation D-3), for reporting. Nothing has been tuned on it.
• Frozen records are not edited. This covers preregistration/, backtests/results/v2*/ and backtest_snapshot/. Re-runs write to backtests/rerun/ and are compared file by file.
• No licensed data in git. This covers Databento and Yahoo prices, and any file derived from them that carries price levels.