Overview
Binance account
JEPA-based predictions
The real JEPA latent-representation model validated an information coefficient of 0.13–0.44 for realized-volatility forecasting in backtest (see project notes) — direction forecasting was tested and refuted. This panel illustrates what a live version could show: a predicted-vs-realized volatility forecast per asset, and a current regime read. The lines below are generated from the synthetic series, not a live model.
Risk management
Monte Carlo simulation
Geometric Brownian motion, parameters fit from the synthetic asset's own historical drift/volatility. Genuinely simulated client-side (not hardcoded) — change the inputs and re-run.
Value at Risk
1-day VaR/CVaR computed two ways from the same synthetic return history: historical (empirical quantile) and parametric (assumes normal returns) — shown together because they typically disagree most exactly when it matters, in the tails.
Suggested additional metrics
Beyond VaR — these catch things VaR alone misses: how bad the average bad day is (CVaR, above), whether the distribution is fat-tailed (kurtosis), how concentrated the book is (Herfindahl), and downside-only risk-adjusted return (Sortino).
| Asset | Skewness | Excess Kurtosis | Sharpe | Sortino | Max DD |
|---|
Spillover: VAR(1) + impulse response
A reduced-form vector autoregression (1 lag) is fit by least squares on the 5-asset synthetic return history. Pick a "shock" asset and see how a 1-standard-deviation move propagates to the rest of the portfolio over the following periods. This is a genuine (if simplified) VAR/IRF calculation, not looked up — but it's a reduced-form shock, not a properly identified structural one (no Cholesky/sign-restriction ordering), so read direction and relative magnitude, not precise causal size.
Volatility analysis & jump detection
Rolling realized volatility (20d, annualized)
Price with flagged jump days
Jump analysis
Two methods, shown together: a simple threshold rule (flag a day if |return| exceeds 4× its trailing 20-day volatility — the red dots on the chart), and a realized-variance decomposition (Barndorff-Nielsen & Shephard bipower variation) that separates total variance into a continuous component and a jump component without needing a threshold at all.