MQFS — Sentiment-Statistical Arbitrage Engine

Mediterranean Quantitative Finance Society
Founder & Lead Researcher: Nicolò Angileri

pairs trading engine that fuses classical statistical arbitrage (Engle-Granger cointegration + rolling z-score mean-reversion) with real-time FinBERT NLP sentiment scored on financial news headlines fetched from public RSS feeds — no API key required.

Trading thresholds are dynamically adjusted based on the directional strength of market sentiment: the engine becomes more aggressive when the market has a clear view and more conservative during noisy, neutral periods.

---

Backtest: KO/PEP vs XOM/CVX

The engine is benchmarked on two classic cointegrated pairs from different sectors:

Pair	Sector	Shared Drivers
KO / PEP	Consumer Staples / Beverages	Input costs, consumer spending, FX, advertising cycles
XOM / CVX	Energy	Crude oil price, refinery margins, capex cycles, OPEC policy

Comparing two pairs from different sectors demonstrates that the engine works generically and is not fitted to a single case.

To reproduce the backtest:

python run_backtest.py

This downloads real prices via yfinance, fetches RSS headlines, runs FinBERT, and saves backtest_results.png.

---

Theory

1. Cointegration (Engle-Granger, 1987)

Two assets are cointegrated if a linear combination of their prices is stationary (I(0)). We estimate the hedge ratio β via OLS and verify stationarity with the Augmented Dickey-Fuller test:

Spread_t = Y_t − β · X_t

ADF test: H0 = unit root (non-stationary), reject if p-value < 0.05

2. Dynamic Z-Score

Z_t = (Spread_t − μ_rolling) / σ_rolling

Z >  +threshold  →  spread overvalued   →  SHORT spread (short Y, long X)
Z <  −threshold  →  spread undervalued  →  LONG  spread (long Y, short X)
|Z| < exit_z     →  mean-reversion done →  EXIT position

3. FinBERT Sentiment Adjustment

News headlines are fetched from Yahoo Finance RSS (free, no API key) and scored with ProsusAI/finbert, a BERT model fine-tuned on financial corpora:

article_score = P(positive) − P(negative)    ∈ [−1.0, +1.0]
daily_score   = mean(article_scores on date)

The absolute magnitude of the normalised daily sentiment lowers the z-score threshold:

adjusted_threshold = base_z − clip(|Z_sentiment| × scaler, 0, cap)

Economic rationale: strong directional sentiment (whether positive or negative) indicates the market has a clear view, making mean-reversion trades more likely to complete. Neutral sentiment leaves the threshold unchanged at base_z.

4. Exit Signal

The position is closed when |Z| < exit_z_threshold (mean-reversion completed). An opposing entry signal forces an immediate exit; the new position opens on the next bar (no same-bar flip — eliminates look-ahead bias).

5. Performance Metrics

The Sharpe ratio is computed on daily mark-to-market returns (not per-trade PnL), then annualised with √252. This is the only dimensionally correct method. Positions still open at end of series are closed at the last available spread price and included in metrics.

---

Architecture

MQFS-Semantic-Arbitrage/
├── mqfs_engine.py          Core: cointegration, signal generation, performance
├── sentiment_provider.py   FinBERT + Yahoo Finance RSS + yfinance fallback
├── run_backtest.py         End-to-end two-pair backtest + 6-panel chart
├── requirements.txt        Python dependencies
└── README.md

---

Installation

git clone https://github.com/nicoloangileri/MQFS-Semantic-Arbitrage.git
cd MQFS-Semantic-Arbitrage
pip install -r requirements.txt

FinBERT requires PyTorch. CPU-only install (recommended unless you have a GPU):

pip install torch --index-url https://download.pytorch.org/whl/cpu

Prefer the lightweight VADER fallback? No GPU or torch required:

pip install vaderSentiment
python run_backtest.py --no-finbert

---

Usage

Run the two-pair backtest (recommended)

python run_backtest.py              # FinBERT, 2-year window
python run_backtest.py --period 1y  # Shorter window
python run_backtest.py --no-finbert # VADER fallback

Produces:

• backtest_results.png — 6-panel chart (2 pairs × 3 panels)
• Full metrics printed in terminal

Programmatic usage

import pandas as pd
from mqfs_engine import SemanticStatArb
from sentiment_provider import SentimentProvider

# 1. Load price series (pd.Series with DatetimeIndex, adjusted close)
asset_x = ...  # e.g. KO
asset_y = ...  # e.g. PEP

# 2. Build sentiment vector from RSS + FinBERT
provider      = SentimentProvider(use_finbert=True)
sentiment_vec = provider.build_daily_sentiment("KO", "PEP", asset_x.index)

# 3. Run the engine
engine  = SemanticStatArb(asset_x, asset_y, lookback_window=20)
coint   = engine.calculate_cointegration()

if coint.is_cointegrated:
    signals = engine.generate_sentiment_adjusted_signals(sentiment_vec)
    metrics = engine.compute_performance_metrics(signals)
    print(f"Sharpe: {metrics.sharpe_ratio:.2f} | WinRate: {metrics.win_rate:.1%}")

---

Example Output

========================================================================
  METRIC                            KO / PEP          XOM / CVX
========================================================================
  Cointegrated                           YES                YES
  ADF p-value                         0.0041             0.0089
  Hedge Ratio (beta)                  1.2345             0.8761
  R^2 OLS                             0.9312             0.9087
  Sharpe Ratio                        1.4200             1.1830
  Max Drawdown                       -0.8134            -1.2045
  Win Rate                            62.50%             58.33%
  Profit Factor                        1.870              1.540
  Total Trades                            16                 18
========================================================================

Note: actual results vary with market conditions and RSS news coverage.

---

News Sources

Source	Type	History	API Key
Yahoo Finance RSS	Primary	Historical articles	None required
yfinance .news	Fallback	Recent only (~30-100 articles)	None required

The fallback activates automatically when the RSS feed is unreachable. Coverage is always reported in logs (e.g. "Sentiment coverage: 47/504 days with real news (9.3%)") so users are never misled about data quality.

---

Key Design Decisions

Decision	Rationale
FinBERT over VADER	Fine-tuned on financial text; measurably superior on domain-specific language (Araci, 2019)
RSS over yfinance.news	Historical coverage; higher article count; no rate limiting
Sentiment as threshold modifier, not signal generator	Avoids spurious signals; sentiment modulates risk appetite, not direction
	Z_sentiment
Exit at	Z
No same-bar flip	Prevents look-ahead bias; position reversal requires one full bar
Sharpe on daily MTM, not per-trade PnL	Dimensionally correct annualisation with √252
Open position at end of series closed MTM	Avoids silent exclusion of incomplete trades

---

Limitations & Future Work

• News coverage: RSS feeds provide recent articles. For deep historical backtests, a paid news API (Bloomberg, Refinitiv, NewsAPI Pro) would give full coverage.
• No transaction costs: Set transaction_cost_per_trade in compute_performance_metrics() to model realistic friction.
• Static hedge ratio: OLS β is estimated once on the full sample. A rolling or Kalman-filter-based dynamic hedge ratio would better handle structural breaks.
• Position sizing: The engine generates binary signals. Kelly criterion or volatility-targeted sizing should be applied before live deployment.
• Single-leg PnL: PnL is in “spread units”. Real deployment requires mapping to dollar P&L via position size and notional value.

---

Academic References

1. Gatev, E., Goetzmann, W. N., & Rouwenhorst, K. G. (2006). Pairs Trading: Performance of a Relative-Value Arbitrage Rule. Review of Financial Studies, 19(3), 797–827.
2. Krauss, C. (2017). Statistical Arbitrage Pairs Trading Strategies: Review and Outlook. Journal of Economic Surveys, 31(2), 513–545.
3. Araci, D. (2019). FinBERT: Financial Sentiment Analysis with Pre-trained Language Models. arXiv:1908.10063.
4. Engle, R. F., & Granger, C. W. J. (1987). Co-Integration and Error Correction: Representation, Estimation, and Testing. Econometrica, 55(2), 251–276.
5. Vidyamurthy, G. (2004). Pairs Trading: Quantitative Methods and Analysis. Wiley Finance.

---

---

MQFS — Mediterranean Quantitative Finance Society