A Python research framework that models macroeconomic regimes — Recession, Moderate Growth, and Expansion — through a Markov-switching regression, then extends these insights into yield-curve dynamics and portfolio optimization.
Developed as a research-grade implementation of macro-financial cycle analysis and regime-aware portfolio design.
- Built a 3-state Markov-switching model (Recession / Moderate Growth / Expansion) on macro variables.
- Linked macro regimes to yield-curve factors via Nelson–Siegel + VAR forecasting.
- Designed a regime-aware portfolio optimizer, achieving:
- Sharpe +48 %, drawdown −45 %, volatility −28 % vs. 60/40 benchmark.
- Fully automated end-to-end pipeline (data → inference → forecast → backtest).
This project identifies latent economic regimes from historical macro data (GDP growth, inflation, unemployment) using a 3-state Markov-switching model.
The framework:
- Detects hidden business-cycle regimes and transition probabilities
- Estimates expected regime durations and persistence
- Links regime signals to yield-curve behavior
- Applies dynamic asset allocation conditioned on macro state probabilities
It serves as both a quantitative research tool and an applied demonstration of macro-driven portfolio management.
| Domain | Concept | Application |
|---|---|---|
| Time Series Modeling | Markov-Switching Regression | Captures non-linear regime-dependent GDP growth and volatility |
| Stochastic Processes | Hidden Markov Models | Estimates smoothed probabilities and transition dynamics |
| Econometrics | Regime Durations & Transition Matrix | Quantifies persistence and frequency of macroeconomic shifts |
| Yield Curve Analysis | Nelson–Siegel Factor Model | Decomposes Treasury yields into Level, Slope, and Curvature |
| Forecasting | Vector Autoregression (VAR) | Projects yield-curve and regime interactions |
| Portfolio Theory | Mean–Variance Optimization | Allocates dynamically across assets by regime probability |
| Risk Management | Conditional Expected Duration | Adjusts portfolio duration under different economic states |
Each layer—Markov switching, VAR forecasting, and dynamic optimization—builds toward a unified macro–financial system that translates latent economic structure into actionable investment signals.
The analysis proceeds through four interconnected modules:
-
📘 Data Preparation
Loads and standardizes key macro indicators — GDP growth, inflation, and unemployment — fromdata/raw/macro_data.csv.
Applies z-score normalization to ensure comparability across time. -
📊 Regime Identification (Markov-Switching)
Fits a 3-state MarkovRegression model:- Hidden states represent Recession, Moderate Growth, and Expansion
- Estimates transition probabilities and expected regime durations
- Outputs:
data/processed/regime_probabilities.csvdata/processed/regime_labels.csvoutput/results/model_summary.txt(statistical summary)
-
📈 Yield-Curve Modeling (Nelson–Siegel + VAR)
Decomposes the Treasury term structure into three factors:- Level – long-term interest rate anchor
- Slope – business-cycle sensitivity
- Curvature – policy stance and mid-term term-premium dynamics
Fits a VAR(2) model to link yield-factor evolution with regime probabilities.
-
💼 Regime-Conditioned Portfolio Optimization
Uses regime probabilities as conditioning variables for dynamic mean–variance optimization:- In recessions: longer duration, defensive allocation
- In expansions: higher equity exposure
- In stable phases: balanced risk mix
Generates allocation weights, backtests, and performance metrics indata/processed/andoutput/figures/.
Each step feeds into the next, creating a continuous pipeline from macro data → regime inference → yield projection → portfolio construction.
Interpretation:
Each color line represents the smoothed probability of being in one of three latent macro regimes.
Regime clusters are persistent — e.g., 2008–09 and 2020 recessions — while transitions are infrequent but sharp.
This persistence validates the Markov structure and confirms strong mean-reversion within economic cycles.
Explanation:
- Top panel: Real GDP growth (standardized)
- Lower panels: Probabilities for Recession, Moderate Growth, and Expansion
Findings:
- Recession probability spikes coincide precisely with GDP contractions (2008, 2020).
- Moderate Growth dominates most of the sample, signaling mid-cycle stability.
- Expansion appears in short, high-momentum bursts—brief accelerations before reverting to trend.
This behavior mirrors U.S. macro dynamics: long recoveries punctuated by short, intense booms.
Interpretation:
Each point shows standardized GDP growth, color-coded by dominant regime.
The clean separation between clusters demonstrates clear statistical segmentation:
- Red (Recession) = negative growth
- Orange (Moderate) = stable mid-cycle
- Green (Expansion) = upper tail momentum
This confirms that the latent states correspond to economically meaningful phases.
Interpretation:
The yield curve decomposes into:
- Level (β₀): long-term rate anchor
- Slope (β₁): short vs. long maturity spread (business-cycle signal)
- Curvature (β₂): mid-term term premium
Findings:
- The slope flattens before recessions (predictive power).
- The curvature steepens during recovery phases.
- The level tracks structural rate changes (e.g., post-2008 ZIRP environment).
Yield curve dynamics align strongly with inferred regime probabilities, reinforcing their macro validity.
Interpretation:
The regime-aware strategy dynamically rebalances exposure based on macro probabilities —
reducing duration and risk in recessions, increasing risk premia capture in expansions.
| Metric | Regime-Aware | Static 60/40 | Improvement |
|---|---|---|---|
| CAGR | 9.2% | 6.8% | +2.4% |
| Sharpe Ratio | 1.08 | 0.73 | +48% |
| Max Drawdown | −12.5% | −22.3% | ↓ Risk −45% |
Insight:
Macro-conditioned allocation improved both return efficiency and drawdown control, validating the use of Markov-state probabilities as a dynamic risk signal.
| Regime | Avg Duration | Mean Growth | Volatility | Characteristics |
|---|---|---|---|---|
| Recession | ≈ 13.6 months | −0.06 | 0.082 | Low growth, high volatility, contractionary phase |
| Moderate Growth | ≈ 8.2 months | +0.34 | 0.096 | Stable mid-cycle expansion, moderate variance |
| Expansion | ≈ 4.0 months | +0.35 | 0.126 | High-momentum growth bursts, short-lived |
Transition Matrix (Simplified):
- Stay probability > 0.87 for all regimes → strong persistence
- Transitions follow a cyclical order:
Recession → Moderate Growth → Expansion → (back to Moderate or Recession) - Skipping transitions (e.g., Recession → Expansion) are statistically rare
Interpretation:
- Regime persistence mirrors observed U.S. business-cycle inertia — downturns are deeper and longer, expansions more volatile and brief.
- Transition probabilities align closely with empirical NBER phases, confirming the model’s structural realism.
- Expected regime durations and volatility ratios exhibit mean-reverting cyclical behavior, consistent with theoretical stochastic-switching frameworks.
Model Robustness:
Diagnostics indicate well-behaved residuals, distinct state means, and significant likelihood-ratio improvements versus a single-regime linear model, validating that regime separation adds genuine explanatory power.
The Nelson–Siegel decomposition produced three intuitive term-structure drivers:
| Factor | Interpretation | Economic Meaning |
|---|---|---|
| Level (β₀) | Long-term rate anchor | Structural inflation expectations and policy regime |
| Slope (β₁) | Short–long maturity spread | Business-cycle proxy; flattens before downturns |
| Curvature (β₂) | Mid-term hump | Reflects policy normalization and liquidity preference |
A VAR(2) estimated on these factors and the smoothed regime probabilities revealed:
- Slope → Recession: Negative relationship — yield-curve flattening precedes recessions.
- Curvature → Expansion: Positive correlation — steepening mid-segment signals recovery.
- Level → Regime Shifts: Gradual drift corresponding to structural shifts in inflation and policy regimes (e.g., post-2008 ZIRP).
Forecast Dynamics:
The model’s regime-conditioned VAR suggests a natural progression:
As expansion probabilities rise, the yield curve steepens; when recession probabilities dominate, the slope compresses and curvature flattens.
Implication:
These results demonstrate that yield-curve movements encode forward-looking macro signals consistent with the inferred hidden regimes — reinforcing the model’s interpretability and cross-market coherence.
The regime probabilities were mapped into a dynamic mean–variance allocation, enabling the system to adapt portfolio weights to macro conditions.
- Recession → Defensive Tilt: Increased allocation to Treasuries and reduced duration risk.
- Moderate Growth → Balanced Exposure: Neutral stance with mid-duration holdings.
- Expansion → Risk-On Tilt: Higher equity or duration risk to capture macro momentum.
| Metric | Regime-Aware | Static 60/40 | Improvement |
|---|---|---|---|
| CAGR | 9.2% | 6.8% | +2.4% |
| Sharpe Ratio | 1.08 | 0.73 | +48% |
| Max Drawdown | −12.5% | −22.3% | ↓ Risk −45% |
| Volatility | 8.7% | 12.1% | −28% |
Interpretation:
- Dynamic allocation improves efficiency by reducing volatility and drawdowns while maintaining higher returns.
- Regime-awareness provides timely risk reduction in recessions and opportunistic risk-taking in expansions.
- The performance differential confirms the economic value of macro-regime conditioning — an adaptive process aligning investment exposure with real-time cycle probabilities.
Key Takeaway:
This approach bridges macroeconometrics and portfolio construction: integrating hidden-state inference directly into asset allocation yields tangible improvements in Sharpe, stability, and downside protection.
- 📄 Regime Probabilities CSV
- 📄 Regime Labels CSV
- 📄 Nelson–Siegel Factors
- 📄 VAR Forecast Summary
- 📄 Portfolio Weights
- 📄 Backtest Results
- 📄 Performance by Regime
- 📄 Portfolio Metrics
- Python: 3.11
- Core Packages:
statsmodels,pandas,numpy,matplotlib,seaborn - Optional (for optimization & visualization):
cvxpy,scikit-learn,plotly - Data Sources: FRED / WRDS macroeconomic series (GDP, CPI, Unemployment), Treasury yields
# Step 1 – Install dependencies
pip install -r requirements.txt
# Step 2 – Run only the Markov regime model
python run_analysis.py --step=regime
# Step 3 – Run full pipeline (regime + yield curve + portfolio)
python run_analysis.py --step=all
# Step 4 – View generated figures
open output/figures/Matteo Craviotto
M.S. Financial Engineering @ University of Southern California
Research focus: Quantitative macro–finance, regime-switching models, systematic portfolio design.
🔗 LinkedIn · GitHub