RaSenTra: Regime-Aware Sentiment Trading Decision Support System

This repository contains the source code and experimental results for the paper "RaSenTra: A Regime-Aware Sentiment-Driven Decision Support System for Risk-Controlled Equity Trading", submitted to Expert Systems With Applications.

Note on anonymity. This repository is shared anonymously for double-blind peer review. Author names, affiliations, and identifying information have been removed and will be added after acceptance.

Overview

RaSenTra converts financial news sentiment into risk-controlled equity trading decisions through three integrated modules:

1. Context-aware sentiment module (src/sentiment.py) — derives sentiment from financial news with FinBERT and aggregates it using confidence, dispersion, and temporal-decay weighting (Eqs. 1–4).
2. Market-regime detection module (src/regime.py) — fits a Gaussian HMM to market features and adaptively reweights the sentiment signal per regime (Eq. 5).
3. Decision module (src/decision.py) — maps the modulated signal to position sizing, adaptive stop-loss, and exposure limits (Eq. 6).

A causal backtest (src/backtest.py) evaluates the strategy net of transaction costs and slippage, reporting Sharpe, Sortino, Calmar, maximum drawdown, and turnover (Eq. 7).

Repository structure

rasentra/
├── README.md
├── requirements.txt
├── LICENSE
├── configs/
│   └── default.yaml         # experiment hyperparameters
├── src/
│   ├── sentiment.py         # context-aware sentiment aggregation (Eqs. 1–4)
│   ├── regime.py            # HMM-based regime detection (Eq. 5)
│   ├── decision.py          # position sizing & risk control (Eq. 6)
│   ├── backtest.py          # causal backtest engine (Eq. 7)
│   ├── metrics.py           # performance metrics
│   └── make_figures.py      # reproduce paper figures from results
├── figures/                 # generated figures
└── results/                 # metrics_fast.json, ablation outputs

Installation

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

Tested with Python 3.10+.

Data

This study uses the FNSPID dataset (publicly available; Dong et al., 2024), which couples S&P 500 financial news with aligned price data. Price OHLCV data are obtained via yfinance. The dataset is not redistributed here; see the paper for retrieval instructions.

Usage

# 1. Compute aggregated sentiment signals
python src/sentiment.py --config configs/default.yaml

# 2. Detect market regimes
python src/regime.py --config configs/default.yaml

# 3. Run the decision module + causal backtest
python src/backtest.py --config configs/default.yaml

# 4. Reproduce paper figures from stored results
python src/make_figures.py --results results/metrics_fast.json --out figures/

Reproducing paper results

The numerical results reported in the paper are stored in results/metrics_fast.json, which contains performance metrics for RaSenTra and three baselines (BuyHold, PriceOnly, SentimentOnly) over the 501-day evaluation window, plus a paired significance test of RaSenTra against the sentiment-only ablation. Running src/make_figures.py regenerates all figures directly from this file without re-executing the full backtest:

python src/make_figures.py --results results/metrics_fast.json --out figures/

License

Released under the MIT License (see LICENSE).