Volatility Signals vs Price Forecasts: What Commodity Traders Actually Need and Why They Are Different

The Directional Forecast Problem

Most commodity price forecasts attempt to answer a single question: will spot be higher or lower at some future date? The modelling literature on directional commodity prediction is vast and the results are, charitably, inconsistent. Point forecasts for metals like copper, nickel, or lithium over 7- to 30-day horizons routinely fail to outperform a random walk once transaction costs and realistic execution assumptions are applied. This is well-documented across academic and practitioner research.

The reason is structural. Commodity prices are driven by a dense, nonlinear interaction of supply shocks, inventory draws, FX moves, speculative positioning, and geopolitical disruptions. The signal-to-noise ratio for directional moves at short horizons is extremely low. A model that achieves 53% directional accuracy on nickel over a 14-day window is not actionable for most systematic strategies after slippage and carry costs.

Volatility, by contrast, clusters. It exhibits serial correlation, responds to identifiable macro and microstructural catalysts, and is measurable in realized terms against which a model can be rigorously scored. A volatility probability forecast does not need to know whether nickel moves up or down. It needs to know whether the magnitude of the move will exceed a given threshold. That is a fundamentally easier statistical problem, and the distinction matters for anyone managing a book with convexity exposure.

Why Vol Regimes Are Actionable Where Price Targets Are Not

Consider the practical use cases. An options desk pricing 30-day straddles on LME copper needs an informed view on realized vol relative to implied. A risk manager setting VaR limits on a concentrated nickel position needs to know whether the current regime is likely to persist or shift. A procurement team hedging aluminium input costs over the next quarter needs to decide between vanilla forwards and option structures based on whether tail moves are plausible.

None of these decisions require a directional price forecast. All of them require a calibrated view on volatility regime. The question is not "where is copper going?" but "how much could copper move, and what is the probability of an outsized move within my risk horizon?"

This is the design principle behind the Volterra model. Rather than predicting price direction, the system classifies the probability of entering one of five volatility regimes (LOW through EXTREME) at 7-day, 14-day, and 30-day horizons. The XGBoost classifier, walk-forward cross-validated with a mean AUC of 0.815, ingests 96 daily GDELT GKG news files alongside geographic concentration indices (HHI), supply chain disruption signals, and market context features. The output is a probability distribution across regimes, not a point estimate.

Asymmetric Payoffs Demand Probabilistic Framing

Directional forecasts implicitly assume symmetric loss functions. You are either right or wrong by some magnitude. But real commodity portfolios carry asymmetric exposures. A short gamma position on tin does not care whether tin rallies 2% or falls 2%. It cares whether tin moves 6% in either direction. A long straddle on cobalt is indifferent to sign but acutely sensitive to the probability of realized vol exceeding implied.

Probabilistic volatility signals map directly onto these payoff structures. When the Volterra pipeline flags ELEVATED or HIGH probability for lithium at the 14-day horizon, that output can be compared against the prevailing implied vol surface to identify mispricing. When the signal is LOW, it informs a different set of trades: short vol positions, calendar spreads harvesting roll-down, or reduced hedge ratios for physical books.

The Volterra dataset captures these probability distributions daily across 12 exchange-traded critical minerals on LME, COMEX, NYMEX, and SGX. Figures from the Volterra daily pipeline. Full historical backfill available on AWS Data Exchange.

Scoring the Right Objective

Model evaluation reinforces the distinction. A price forecast is scored on RMSE or directional accuracy, metrics that penalize every miss equally regardless of market regime. A volatility classifier is scored on AUC, calibration curves, and regime-conditional accuracy, metrics that reward the model for correctly identifying the transitions that matter most to risk-adjusted PnL.

For desks running systematic vol strategies or risk teams calibrating scenario frameworks, the relevant question is not whether a model can beat the forward curve on direction. It is whether the model reliably distinguishes between regimes where realized vol will be 15% annualized versus 40% annualized. That classification, delivered with calibrated probabilities and transparent feature attribution, is what converts model output into positioning decisions.