Project Results

Regime Identification

We considered three common time-series clustering algorithms:

  1. Dynamic Time Warping (DTW)

  2. Soft Dynamic Time Warping (S-DTW)

  3. Global Alignment Kernel with K-Means (GAK-KM)

The graphs below show all per-regime (cluster) returns on the left, with their cumulative return on the right. It's evident that GAK-KM was best-suited for this task since:

  • The Regime sizes are well-distributed

  • The Regime returns are interpretable for all macro assets

Dynamic Time Warping Outputs

Regime Sizes: 100, 641, 1241, 308

Volatility Index (VIX)

Three out of four regimes are parabolic, while blue exhibits more volatility.

Dollar Index (USD)

Green relatively bullish. Other regimes exhibit high volatility

Oil (OIL)

Blue and Red clearly express Bear and Bull regimes with parabolic cumulative returns, while Green and Yellow exhibit more volatility. Green is significantly bearish, while Yellow's returns are close to flat.

Soft Dynamic Time Warping Outputs

Regime Sizes: 10, 428, 394, 1458

Dollar Index (USD)

All regimes exhibit volatility, with only Green showing a deliberate trend.

Volatility Index (VIX)

Three out of four regimes are parabolic. The most frequent regime, Red, exhibits a strong bullish trend, cumulatively returning 200%.

Oil (OIL)

We see parabolic behavior from all regimes but Red, which returns flat after 1458 trading days.

GAK K-Means Outputs

Regime Sizes: 214, 282, 835, 959

Volatility Index (VIX)

We see clear bearish and bullish trends, with parabolic and measured regimes for each.

Dollar Index (USD)

All regimes express volatility, and all but red show a consistent trend.

Oil (OIL)

All four regimes show clear trends.

Regime Classification

We trained an XGBoost classifier on the data (Price_t, Regime_t) corresponding to the above MacroEconomic assets, as well as several MicroEconomic assets. We achieved a training accuracy of 98.985%, and a testing accuracy of 92.979%. This shows us that the XGBoost model is able to recognize the pattern in the underlying regimes, signaling the viability of our analysis for applications in trading strategies.

Each XBGoost feature's F-Score is shown on the left. All input features are listed below.

  • f2: OIL

  • f0: VIX

  • f1: Dollar Index

  • f6: 1-Yr US Treasury Bond

  • f3: Corporate Credit Spread

  • f9: TIPS

  • f5: ·periodspread

  • f7: 2-Yr US Treasury Bond

  • f8: 10-Yr US Treasury Bond

  • f4: Monetary Base