Commodity Price Forecasting
GAMLSS · Bayesian MSGARCH · Rolling-origin validation · R
The business question
A grain buyer does not only need to know where the price is heading. They need to know how wrong that number could be, and whether being wrong costs the same in both directions.
Most forecasting stops at the first question. A cooperative deciding when to sell, or a processor deciding how much to hedge, is really asking the second one.
The data
Public series only: agricultural commodity prices from international exchanges and Brazilian domestic markets, daily, over several years, with the exchange rate as the bridge between them. No client data is involved at any point.
Method, and why this one
The usual simplification. Standard volatility models let the spread of returns move over time but hold everything else fixed — the asymmetry and the tail weight are treated as constants. In agricultural markets that assumption is uncomfortable: a drought does not widen the distribution symmetrically.
Letting the whole distribution move. GAMLSS gives each parameter of the distribution — location, scale, skewness, kurtosis — its own model with its own covariates. The shape of the distribution becomes something estimated over time rather than assumed.
Regimes, not gradual drift. Grain markets do not slide smoothly from calm to turbulent; they switch. Markov-switching GARCH estimates the states and the transition probabilities between them, and a Bayesian treatment carries the parameter uncertainty forward instead of collapsing it into a point estimate.
Validation that respects time. A model tuned on the full sample will always look good. Rolling-origin evaluation refits at each step using only what was knowable at that moment, and every added layer has to earn its place against simpler benchmarks — a plain GARCH, an ARIMA, a random walk. A more elaborate model that does not beat the simple one is a finding, not a failure.
Result
Deliberately not published here. The manuscripts behind this work are under peer review, and putting their results on a personal website before the journals have ruled on them would be poor practice — and unfair to the reviewers reading them.
What can be said is the shape of the work: this is one stage of a doctoral thesis on multi-period, multi-objective portfolio optimisation in agricultural commodity markets, where distributional modelling feeds regime-switching volatility, which feeds scenario simulation, which feeds the optimisation layer. Each stage stands as its own manuscript.
Whatever of this line reaches print is listed on the research page, alongside the conference papers that already carry proceedings links. The case below is the part that is public and running today.
Code
The model ladder described here — MSGARCH first, backing down to GARCH and then ARIMA when the residual diagnostics reject it — is implemented and running in the multi-agent case study, with public data and a frozen cache so anyone can reproduce it:
Multi-Agent Commodity Analyst Repository
Published work along this line is listed on the research page, where four conference papers carry links to their proceedings.