🚨 Black Swan Risk Indicator

Early-warning system for financial market crises · NLP sentiment + ML forecasting + regime detection

📅 Last updated: 2026-05-28 09:12 UTC 📊 5 years of data View source on GitHub →

🌤️ Today's Market Snapshot

30-second overview — key indicators investors watch + regime interpretation

☀️

Low Risk — Markets in Normal Mode

Low VIX + Trending Bull regime → standard investment strategies remain viable

🌡️

Current VIX

16.71

☀️

🎭

Market Regime

📈 Trending Bull

Uptrend — SMA-50 > SMA-200 and low volatility

📊

S&P 500

7,520

+4.32% (20 day)

🌡️ VIX — The Market Fear Index

CBOE Volatility Index — derived from S&P 500 option implied volatility, 30-day horizon

VIX < 15 = very calm (possibly complacent)
VIX 15-20 = normal bull market range
VIX 20-30 = investors getting nervous
VIX 30-40 = crisis signal (2018 rate hikes, 2022 inflation)
VIX > 40 = full crisis — all-time peak: 82.69 (Mar 16, 2020 COVID)

💬 💡 How to read: Blue line = actual VIX · Dotted orange = 20-day moving average · Red star = XGBoost's prediction for VIX 7 trading days from now

📈 VIX history + AI forecast (7-day ahead)

💡 How to read: Blue line = actual VIX · Dotted orange = 20-day moving average · Red star = XGBoost's prediction for VIX 7 trading days from now

🔮 ML Forecast: 7-Day VIX Prediction

XGBoost trained on VIX lag features + macro indicators (yield curve, gold, oil, DXY) — but check accuracy in the next section before trusting

📈

Forecast: VIX rising

Model expects VIX at 18.8 (delta +2.1 from today) — heightened risk monitoring advised

💬 Important caveat: VIX is notoriously difficult to forecast — even academic papers regularly report R² near zero. Treat this as a directional signal, not a precise number. See accuracy section below for the honest assessment.

🎯 But... How Accurate Is It?

Evaluated with walk-forward cross-validation (TimeSeriesSplit) — the gold standard for time-series ML that prevents look-ahead bias

R² (R-squared) = how much better the model explains variance vs predicting the mean
R² = 1.0 = perfect prediction (never happens in practice)
R² = 0 = same as predicting the mean — useless
R² negative = worse than predicting the mean — model is harmful

💬 Model cannot beat the mean prediction — consistent with the Random Walk Hypothesis (Fama, Nobel 2013). Markets contain enough randomness that ML often underperforms simple baselines on volatility forecasting.

⚔️ Model Comparison

Tested 4 ML models + 1 Naive baseline with walk-forward CV — who wins at VIX forecasting?

👶 Naive baseline: persistence model — predicts "VIX in 7d = VIX today" (no ML)
🌲 XGBoost: gradient-boosted trees, regularized (max_depth=3, n_estimators=100)
💡 LightGBM: fast tree-based with leaf-wise growth
📐 Ridge: linear regression with L2 regularization
🧠 LSTM: 1-layer PyTorch LSTM (32 hidden units, dropout 0.2)

🥇

Naive

R² = 0.101

Highest R² in this evaluation

🥉

LSTM

R² = -3.744

Lowest R² in this round

💬 Surprise! Naive baseline beats every ML model

The "VIX in 7d = VIX today" model (R² = 0.101) beats XGBoost, LightGBM, Ridge, and LSTM. This is the Random Walk Hypothesis in finance — asset prices move stochastically, making short-term prediction extremely difficult.

Engineering lesson: Always test naive baselines first. If your ML doesn't beat persistence, don't deploy it.

🔍 SHAP: What Drives Model Decisions?

SHAP (SHapley Additive exPlanations) uses game theory to attribute predictions to features — the gold standard for interpretable AI

🎯 SHAP values derived from Shapley values (Nobel Economics 1953)
📊 Shows each feature's marginal contribution to a prediction
✅ Fair & consistent — superior to XGBoost's built-in feature_importance
🔍 Useful for debugging model behavior

💬 Model relies most heavily on SP500_Vol_20D — this is the dominant signal driving predictions

🎬 Animated Timeline

Click ▶️ Play to watch VIX evolve month-by-month over 5 years — every crisis is visible (COVID, Russia-Ukraine war, 2022 inflation)

💬 Notice how VIX spikes during every black swan event: Feb 2020 (COVID), Feb 2022 (Russia war), Sep 2022 (Fed rate hikes). This is the recurring pattern we try to detect.

🌐 3D Visualization

3D scatter: time × S&P daily return × VIX — drag with mouse to rotate · scroll to zoom

💬 Observable cluster: red points (high VIX) tend to coincide with days when S&P fell > 3% — confirming the well-known negative correlation between VIX and S&P returns (typically -0.7 to -0.8).

💰 Backtest: Is the Strategy Profitable?

Historical simulation over 5 years — start with $1M, compare "Black Swan Strategy" vs "Buy & Hold" (including 10 bps transaction costs)

🛡️

Black Swan Strategy

1,664,907

Initial $1,000,000 → +66.5%
Max Drawdown: -28.0%

📊

Buy & Hold (baseline)

1,788,811

Initial $1,000,000 → +78.9%
Max Drawdown: -25.4%

💰 Includes 20 trades with 10 bps cost per turnover — realistic conditions

📈 Equity Curve Comparison

📉 Which strategy has shallower drawdown?

Less depth = less pain. A good strategy should have shallower drawdown than the benchmark.

🚨 When did signals fire?

Red triangle markers = days the rule-based detector flagged risk

💬 ❌ Strategy underperforms Buy & Hold in this window — because the past 5 years were a bull market, defensive systems sacrifice upside by sitting in cash during rallies.

The strategy's value emerges during actual crises like COVID-2020 or 2008.

🧪 Paired t-test: Is the strategy significantly better than baseline?

Hypothesis test — H₀: mean(strategy − baseline) ≤ 0 vs H₁: mean(strategy − baseline) > 0 (one-sided) at α = 0.05

t-statistic

-0.5527

p-value (one-sided)

0.7097

α (significance)

0.05

n (daily obs)

1,258

⚠️ Fail to reject H₀ — insufficient evidence that strategy beats baseline (p ≥ α). Consistent with the Random Walk Hypothesis.

🌍 Multi-Asset Volatility

Realized volatility across asset classes: US equity, crypto, gold, emerging markets — if all assets spike together = systemic crisis

💬 Bitcoin's realized volatility is typically 3-5× higher than S&P 500. Gold acts as a "safe haven" — often uncorrelated or negatively correlated with equities during crises.

If all assets spike together = systemic risk signal (e.g., COVID-2020, when even gold fell briefly).

📖 Bottom Line

🌡️ Today's market
☀️ Trending Bull · VIX = 16.71

🔮 7-day forecast
⚠️ VIX expected to rise to 18.8 (+2.1)

💡 What this system does
An early-warning system for market regime changes. Combines 3 signals: news sentiment (FinBERT) + ML forecast + rule-based crisis detector. It does NOT predict the future with certainty — but it helps reduce drawdowns during crises.

⚠️ Disclaimer
This is an educational and research project — not financial advice. Past performance does not guarantee future results. Markets can stay irrational longer than you can stay solvent.

For engineers / data scientists

📊 Technical Numbers

VIX: 16.71
AI Forecast (7d, XGBoost): 18.82
Walk-Forward Mean R²: -0.094 ± 0.571 (5 folds)
Best Model (regression): Naive
Backtest Sharpe: Strategy 0.74 vs B&H 0.77
Max Drawdown: Strategy -28.02% vs B&H -25.43%
Transaction cost: 10 bps per turnover
Features: 13 (VIX lags, S&P returns, yield curve, gold/oil/DXY)

📐 Methodology

Time-series validation: TimeSeriesSplit, no shuffle (no look-ahead bias)
FinBERT (ProsusAI) for news sentiment (separate Streamlit app)
SHAP TreeExplainer for feature attribution
Multi-asset volatility: realized vol 20d annualized (proxy for non-VIX assets)
Regime: SMA-50/200 crossover + 20d rolling vol vs historical median

Full source code: github.com/oomNoNe/black-swan-indicator