Quantitative Research · Portfolio Project

Latent Liquidity & Systemic Risk Dashboard

Bayesian state-space model inferring latent market stress from daily microstructure data. Two particle filter architectures — an adaptive Differentiable PF with Sinkhorn OT resampling and a fixed-parameter Bootstrap PF — are compared across filtering quality and 30-day early-warning performance.

Read the paper → Strategy & PnL → DPF vs BPF Results → Theory & Methods →

DPF · Sinkhorn OT Bootstrap PF SVL + DRD Covariance Crisis Early Warning 9 Historical Episodes

0.567

Endo. Episode AUC

vs BPF 0.500

0.682

Daily ROC-AUC

vs BPF 0.670

340×

Crisis/Calm Ratio

GFC 0.78 · calm 0.002

47–60d

Warning Lead Time

5 of 9 episodes detected

4,526

Training Sample

trading days · 2007–2026

Model Comparison

DPF vs BPF at a glance

DPF

Differentiable Particle Filter

ArchitectureSinkhorn OT resampling (ε=0.5)

ParametersOnline gradient (Adam, η=5e-4)

ParticlesN = 500

Brier Score0.212

Endo. Episode AUC0.567

Signal-to-noise2.67

BPF

Bootstrap Particle Filter

ArchitectureMultinomial resampling

ParametersFixed (hand-calibrated)

ParticlesN = 1000

Posterior CI width0.219

Tail prob. AUC0.726

Mean lead time50.2d

Interactive

Time Window

Select an era or custom range

Start Date

End Date

Loading data…

Signal

60-day Forward Crisis Probability

Logistic early-warning layer · dashed = 0.50 threshold

Rendering…

Recovered Latent Factor

Liquidity Stress L_t with 90% CI

Threshold L=2 · bar = tail prob P(L>2)

Rendering…

Filter Internals

Latent Factors h_t (log-vol) and z_t (correlation driver)

DPF posterior means

Rendering…

Reference Figures

Key Output Figures

Click to enlarge · see results page for full comparison

Performance dashboard — **Performance Dashboard** — DPF vs BPF across all key metrics
`figure_5_performance_dashboard.png`
View High-Res Source

Endogenous vs exogenous — **Endogenous vs Exogenous Breakdown** — why exogenous shocks produce no advance warning
`figure_6_crisis_type_breakdown.png` View High-Res Source

L_t comparison — **L_t Stress Comparison** — DPF vs BPF posterior with 90% CIs
`figure_1_stress_comparison.png` View High-Res Source

GFC animated comparison — **GFC Animation (2007–2009)** — dual-filter forecast cones evolving in real time
`research/animate_comparison.py` View High-Res Source

Brexit animated comparison — **Brexit Window Animation (2016)** — captures early market microstructure dynamics
`research/animate_comparison.py` View High-Res Source

COVID animated comparison — **COVID Window Animation (2019–2021)** — illustrates zero advance warning for exogenous shock
`research/animate_comparison.py` View High-Res Source

Run It Yourself

Reproduce Results

Full results page →

bash

$ pip install -r requirements.txt

$ python main.py run # full pipeline: train → predict → calibrate → predict → bench

# Or run individual steps:

$ python main.py train # learn DPF parameters (rolling 120d windows)

$ python main.py predict --filter dpf # DPF filter + forward simulation

$ python main.py calibrate --filter dpf # fit logistic warning model

$ python main.py bench # DPF vs BPF comparison + all figures

$ python main.py animate --start 2007-06-01 --end 2009-06-01 # animated GIF

Latent Liquidity & Systemic Risk Dashboard

DPF vs BPF at a glance

Time Window

60-day Forward Crisis Probability

Liquidity Stress Lt with 90% CI

Latent Factors ht (log-vol) and zt (correlation driver)

Key Output Figures

Reproduce Results

Liquidity Stress L_t with 90% CI

Latent Factors h_t (log-vol) and z_t (correlation driver)