Real-Time Risk: How AI is Replacing End-of-Day VaR with Intraday Stress Testing

VaR answers a specific question: given my current portfolio positions as of market close, what is the maximum loss I can expect at a given confidence level (typically 95% or 99%) over a given horizon (typically 1 day), based on historical price movements? The limitations embedded in that definition are significant. The 'as of market close' constraint means VaR is calculated on stale positions — a desk that has traded actively during the day has a risk profile that bears no relation to its end-of-day snapshot. The 'based on historical price movements' constraint means VaR systematically underestimates tail risk in novel market regimes — the fat tails that cause actual crises are not well-represented in the historical window used for calibration. And VaR produces a single point estimate when what risk managers actually need is a conditional picture: not just 'what is the 99th percentile loss' but 'what happens to our portfolio in specific stress scenarios — a Fed rate shock, a credit event, a liquidity crisis in digital assets.'

Real-time risk monitoring requires a fundamentally different architecture from batch overnight VaR calculation. Positions must be updated continuously as trades execute, and risk measures must recompute on the updated position set — ideally in under 100 milliseconds for electronic trading desks where market conditions change at millisecond frequency. The technical stack for this requires: a real-time position management system that receives trade confirmations via low-latency feeds and maintains a continuously updated book; a market data feed that prices those positions against current mid-prices across all relevant markets; a risk calculation engine that recomputes exposure metrics (DV01, delta, vega, credit spread sensitivity) on the updated positions; and a limit monitoring system that compares live exposures against pre-approved limits and alerts when thresholds are breached. This architecture exists at tier-1 banks for equities and rates; extending it to cross-asset portfolios that include illiquid assets and OTC derivatives is the frontier.

Scenario stress testing — applying specific market shocks to the portfolio and measuring the P&L impact — has traditionally used a predetermined set of historical scenarios (2008 GFC, 2010 Flash Crash, 2020 COVID sell-off) and regulatory scenarios (Basel III prescribed shocks). The limitation is that history is not a complete scenario space — the next crisis will have features that distinguish it from previous ones, and a predetermined scenario set cannot anticipate them. AI scenario generation uses adversarial approaches to identify the worst-case scenarios for a specific portfolio: given the current position set, what combination of market moves produces the maximum loss? Generative models trained on historical market data can generate novel but historically-plausible scenarios — not just replaying history but extrapolating the dynamics that produce correlated market stress. This allows risk managers to understand their specific portfolio's vulnerability landscape rather than how it performs in generic historical scenarios.

Standard risk measures assume positions can be liquidated at current market prices — an assumption that fails precisely when it matters most, in a crisis. Liquidity-adjusted VaR (LVaR) incorporates the market impact of liquidating a position: a large position in an illiquid instrument cannot be sold at the current mid-price without moving the market against itself. AI-driven liquidity models estimate liquidation cost as a function of position size relative to average daily volume, current bid-ask spread, and market depth — and they can update these estimates in real time as liquidity conditions change. In the 2020 COVID crisis, bid-ask spreads in credit markets widened 5–10× within hours; a risk system that models liquidity as a static constant would have dramatically understated the cost of risk reduction. Intraday liquidity monitoring is the difference between being able to de-risk in an orderly fashion and being forced to liquidate at crisis prices.

Portfolio diversification relies on correlations between asset classes being stable: if equities and bonds have low or negative correlation, holding both reduces portfolio volatility. But correlations are not stable — they spike toward 1.0 in crisis conditions, precisely when diversification benefits are needed most. This is the correlation breakdown problem, and it is the reason that many risk models underestimate drawdown severity in tail scenarios. AI risk models that incorporate regime detection — identifying whether the current market is in a normal, stressed, or crisis regime — and use regime-conditional correlations produce significantly more accurate tail risk estimates than models that assume static correlations. Hidden Markov Models and deep learning approaches to regime detection have both shown strong results in detecting regime transitions before they are apparent from linear statistical measures.

The shift from end-of-day VaR to intraday AI-driven risk management has regulatory implications that institutions must navigate carefully. Basel IV's Fundamental Review of the Trading Book (FRTB) moves risk capital calculation toward Expected Shortfall (ES) rather than VaR, recognising VaR's inadequacy for capturing tail risk. FRTB also increases scrutiny of internal model approval — banks using internal models for capital calculation must demonstrate that those models perform well under backtesting against actual trading outcomes. AI-generated scenarios and regime-conditional correlation models introduce model risk that regulators require to be quantified and managed: the risk that the AI model itself is wrong is a material input to capital calculation. Leading institutions are investing in model risk management frameworks that track AI risk model performance continuously and maintain approved fallback methodologies for each risk calculation.

Found this useful?

Share the key finding on X.