Commodity markets are adversarial to point forecasts. A model that says "haddock will be 280p per kilo next Tuesday" is not just wrong when it misses — it is actively dangerous, because it gives buyers and sellers a false sense of certainty at the moment decisions are made.
The question is not whether a single-point forecast will be wrong. It will be. The question is whether your forecast system communicates the shape of uncertainty in a way that supports better decisions.
The problem with consensus forecasts
Most market intelligence products converge on a consensus view: a central estimate, maybe with a vague range. This is useful for reporting. It is not useful for decisions.
Consider a fish merchant deciding whether to buy at today's auction or wait. The relevant questions are not:
- What is the expected price?
- What is the range of prices?
The relevant questions are:
- What is the probability that prices drop below my margin threshold this week?
- What scenarios would cause an unusual price movement, and how likely are they?
- Given current market structure, is this a regime where my model's assumptions still hold?
A consensus forecast cannot answer these questions. A probabilistic forecast system can.
Regime detection as a prerequisite
Before applying any forecast model, the most important diagnostic question is: which regime are we in?
Commodity markets exhibit distinct structural regimes — periods where the driving variables, their relationships, and their volatility all shift. A model calibrated on a normal-volatility regime will systematically underestimate uncertainty during a disruption regime, precisely when uncertainty matters most.
Our approach to regime detection uses a combination of:
Structural break tests on key price and volume series to identify when the data-generating process has shifted.
Latent state models that infer the current regime from a combination of observed indicators — landing volumes, species mix, seasonal patterns, environmental conditions — without requiring a pre-specified number of regimes.
Volatility regime classification that distinguishes between low-information periods (quiet, trending markets) and high-information periods (reversals, supply shocks) where forecast distributions need to be explicitly wider.
The output is not a single regime label but a probability distribution over regimes. This propagates forward into the forecast.
Constructing forecast distributions
Given a regime probability distribution, we construct calibrated forecast distributions rather than scenario trees with labelled narratives. For markets with daily auction clearing and short decision horizons, the actionable output is a distribution over the next one to twenty trading days — not a 3-month scenario tree.
The distinction matters: conditional distributions are more actionable than point estimates, but they need to be honest about their width. A distribution that says "there is a 20% chance price drops more than 15% this week" is actionable. A range that says "prices will be between 240p and 320p" — without a probability attached — is not.
In practice, we use conformal prediction to produce intervals with guaranteed coverage. If the model says 80% interval, the actual price falls within that interval 80% of the time, measured out of sample. This is not a modelling assumption — it is an empirical guarantee from the calibration method.
Calibration and the track record problem
Probabilistic forecasts are only useful if they are calibrated. A forecast that says "70% confidence" should be right 70% of the time. Most commercial forecasts are not calibrated in this sense; they are constructed to appear confident.
Building a calibrated forecast system requires:
-
Historical back-testing with proper out-of-sample discipline — not just checking that the model fits the historical data, but checking that it would have produced calibrated probabilistic forecasts at each historical decision point.
-
Regular recalibration — market structure changes. A model calibrated on one period may be systematically miscalibrated later if quota regimes, fleet composition, or demand conditions have shifted. Recalibration should be scheduled, not reactive.
-
Honest uncertainty accounting — this is the hardest part. The temptation is always to tighten forecast distributions to appear more useful. Resist it. A wide honest distribution is more useful than a narrow confident one.
What this looks like in practice
In FishFacts, our demersal fish market forecasting tool, this manifests as:
- Day-ahead probabilistic distributions for H1 through H20 (one to twenty trading days ahead), at species and grade level
- Conformal prediction intervals that are empirically validated out of sample — when the model says 80%, it has been right approximately 80% of the time
- Calibration history showing how past forecasts have performed against actuals, organised by species, grade, and horizon
- Benchmark comparison against a naive baseline at every species-grade combination — the system must demonstrate genuine predictive improvement, not just track the market average
The goal is not to replace buyer and seller judgement. It is to give market participants a clearer picture of what the current market structure implies, so that their judgement operates on better inputs.
Uncertainty is not a failure of analysis. It is the honest representation of a genuinely uncertain world. The tools we build are designed to make that uncertainty visible, quantified, and actionable.