Financial Time Series: Volatility Modeling and Risk Forecasting with TSFMs

Financial markets generate time series data at extraordinary scale and complexity. Stock prices, foreign exchange rates, commodity futures, bond yields, and volatility indices like the VIX each carry distinct statistical signatures that have frustrated conventional forecasting methods for decades. The emergence of time series foundation models opens a different angle on these problems, not as a replacement for quantitative trading models, but as a powerful tool for operational risk, treasury forecasting, and regulatory capital estimation.

What Makes Financial Time Series Different

Financial data exhibits a cluster of statistical properties that rarely appear together in other domains. Understanding these properties is essential before applying any model, foundation or otherwise.

Heavy tails. Asset return distributions consistently show fatter tails than the normal distribution predicts. Events that a Gaussian model would place at six or seven standard deviations occur with observable frequency. The 2010 Flash Crash, the 2015 Swiss franc de-peg, and the March 2020 COVID sell-off all produced moves that normal-distribution risk models dramatically underestimated.

Volatility clustering. Large price moves tend to follow large price moves, and calm periods follow calm periods. This autocorrelation of squared returns, first documented by Mandelbrot in 1963, means that risk is not constant over time. A model that assumes stationary variance will systematically underestimate risk during turbulent regimes and overestimate it during quiet ones.

Leverage effects. For equities, negative returns tend to increase future volatility more than positive returns of the same magnitude. This asymmetry, where falling markets become more volatile than rising ones, creates a skew in the conditional return distribution that symmetric models miss.

Regime changes. Markets shift between fundamentally different operating modes: low-volatility trending regimes, high-volatility mean-reverting regimes, and crisis regimes where correlations spike and normal relationships break down. These transitions are abrupt and difficult to predict.

Traditional Approaches and Their Limits

The GARCH (Generalized Autoregressive Conditional Heteroskedasticity) family of models has been the workhorse of volatility modeling since Bollerslev introduced it in 1986. GARCH captures volatility clustering by modeling the conditional variance as a function of past squared returns and past variances. Extensions like EGARCH handle leverage effects, and regime-switching GARCH models attempt to capture structural breaks.

EWMA (Exponentially Weighted Moving Average) volatility, popularized by RiskMetrics, offers a simpler alternative that weights recent observations more heavily. Both approaches share fundamental limitations: they are univariate, parametric, and require explicit specification of the volatility dynamics. When the true data-generating process deviates from the model assumptions, and it always does, forecast quality degrades.

Where TSFMs Add Value

Time series foundation models bring several advantages to financial risk forecasting. Their strength lies not in predicting the direction of markets but in characterizing the distribution of future outcomes.

Probabilistic output for VaR and CVaR. Value at Risk requires estimating specific quantiles of the return distribution. CVaR (Conditional Value at Risk, or Expected Shortfall) requires estimating the expected loss beyond that quantile. Both demand accurate modeling of the distribution's tails, exactly where parametric assumptions tend to fail. Models like Lag-Llama, which output full distributional forecasts rather than point estimates, naturally produce the quantile estimates that VaR and CVaR require. Lag-Llama's Student-t and negative binomial distribution heads are better suited to heavy-tailed financial data than Gaussian assumptions.

Regime detection through anomaly scoring. Abrupt shifts in market behavior, whether flash crashes, liquidity crises, or volatility regime changes, can be detected through the forecast-residual anomaly detection framework. When a TSFM's prediction intervals are systematically violated, it signals that the market has entered a regime the model's recent context does not represent. This is directly useful for triggering risk limit adjustments or hedging overlays.

Zero-shot generalization across instruments. A GARCH model fitted to EUR/USD will not transfer to crude oil futures without refitting. A pretrained TSFM can produce zero-shot forecasts across asset classes because it has learned general temporal patterns from diverse training data. For organizations managing risk across hundreds of instruments, this eliminates the burden of maintaining per-instrument model pipelines.

Regulatory Context: Basel III/IV

Bank capital requirements under Basel III and its successor framework depend directly on risk forecasts. Internal Models Approach (IMA) banks must demonstrate that their VaR and Expected Shortfall models accurately capture tail risk, subject to backtesting by regulators. Underestimating risk leads to capital add-ons; overestimating it ties up capital unnecessarily.

TSFMs offer a practical advantage here: their probabilistic calibration can be validated empirically against prediction interval coverage, providing the kind of backtesting evidence regulators require. The ability to generate forecasts without extensive per-instrument tuning also simplifies the model governance burden that banks face when maintaining hundreds of approved risk models.

What TSFMs Will Not Do

It is important to be direct about boundaries. TSFMs will not generate trading alpha. Profitable trading strategies depend on identifying mispricings that other market participants have missed, and a pretrained model operating on publicly available price data has no information advantage over the market. The efficient market hypothesis imposes hard limits on what any model, foundation or otherwise, can extract from historical prices alone.

Where TSFMs excel is in operational applications where the goal is not to beat the market but to accurately characterize uncertainty: treasury cash flow forecasting, corporate FX exposure management, insurance reserve estimation, and the regulatory capital calculations described above. These are problems where a well-calibrated probabilistic forecast is directly valuable, and where the engineering simplicity of zero-shot inference over maintaining bespoke GARCH pipelines per instrument is a genuine operational advantage.

Getting Started

Financial time series data is widely accessible through sources like Yahoo Finance. You can pull historical price data for any publicly traded instrument, compute returns, and submit the series to TSFM.ai for probabilistic forecasting. Start by exploring the available models in the model catalog and running experiments in the playground. For guidance on moving from experimentation to a reliable production system, see Building Production Forecast Pipelines.