How to Tell Whether Futures Add Information
The question in this chapter is not whether commodity futures prices differ from today’s spot price; they often do. It is whether the difference helps forecast the later spot price after we compare futures with plausible alternatives. The Federal Reserve study makes that question concrete with three benchmarks, a relative error measure, and a sample that varies by commodity and maturity. Understanding that design lets readers judge what a result means—and what it cannot mean—before looking at the reported numbers.
A forecast has no useful meaning without a comparison base. Saying that a one-year futures price was “accurate” could mean that it was close to the later spot price, but it could also mean that it did better than simply carrying today’s price forward. Those are different claims. The first asks about absolute error. The second asks about incremental value relative to a particular method. The study’s benchmarks are designed to make the second question visible.
Three ways to forecast without the futures curve
The first benchmark is a random walk. In this setting, the best forecast of the future spot price is simply today’s spot price. The model makes no claim that prices will rise, fall, or return to an average. It says that, without additional information, the current price is a reasonable starting point for the future. This is a demanding baseline in one sense: a futures price must add information beyond the fact that the market already knows today’s price.
The second is a momentum forecast. It extends recent price changes into the future. If the commodity has risen over the recent lookback period, the momentum forecast expects some continuation; if it has fallen, it extends that movement in the other direction. The Federal Reserve specification describes this as a random walk with drift in logarithms. The important idea is not the notation but the assumption: recent shocks are treated as carrying forward.
That assumption can be reasonable for a while, but it can also fail abruptly. A price move caused by a temporary supply problem may not persist. A structural change may make the recent past a poor guide to the next year. Forces that offset one another can also stop a recent trend from continuing. Momentum is therefore a useful comparator because it asks whether the information in the futures curve is better than simply extrapolating what has just happened.
The third is a mean-reversion forecast. It assumes that recent price changes tend to move back toward a longer-run pattern. In the study’s specification, the long-run average price change is assumed to be zero, and the recent change is adjusted using an estimated negative relationship between current and lagged price changes. The result does not predict an immediate return to a fixed price. It predicts that the latest movement should be moderated when the history suggests that changes tend to reverse.
These are not competing declarations about what the market must do. They are controlled ways to ask what the futures curve contributes. A random walk ignores recent direction. Momentum gives recent direction persistence. Mean reversion gives it a tendency to weaken. Futures prices may contain information about forward-looking supply and demand conditions that none of those backward-looking constructions captures. But that possibility has to be demonstrated against the alternatives, not assumed because futures have a later maturity.
The comparison is meaningful only because the inputs and horizon are defined. The same futures quote could look impressive against a no-change forecast and unremarkable against a stronger model. A relative result also does not state whether the future market was easy or difficult to predict in absolute terms. It answers a narrower question: which of the named methods produced the smaller error under the study’s design?
Relative RMSE is a comparison, not a score out of ten
The Federal Reserve note uses relative root-mean-squared error, or relative RMSE. RMSE summarizes forecast errors by giving larger errors more weight before combining them into one measure. The study then divides the futures RMSE by the RMSE of the selected benchmark. A value below one means that futures had lower error than that benchmark; a value above one means the benchmark did better; a value near one means the difference was small on this measure.
The denominator matters. Relative RMSE is not an absolute grade assigned to the futures market. A ratio of 0.8 does not mean that the futures price was “80 percent correct.” It means that the futures error measure was 0.8 times the error measure of the named comparator. If the comparator changes, the ratio can change even though the futures price and the later spot outcome do not. This is why a single statement that futures “work” is incomplete unless it identifies the baseline.
The measure also rewards a method for being better on average across the observations included in a comparison. It does not imply that every individual futures quote beat the benchmark. A median relative RMSE below one can coexist with episodes in which futures were worse. Nor does a lower RMSE tell us why the method performed better. The mechanism might involve forward-looking information, a benchmark’s poor response to a shock, or the distribution of difficult observations in the sample.
This distinction is central to the difference between information and an unbiased promise. Imagine a futures price that incorporates a real change in expected supply but also reflects the terms on which risk is transferred. It might adjust in a direction that helps a forecaster rank future outcomes, yet remain systematically too high or too low relative to later spot prices. Relative RMSE can detect an advantage over a baseline without proving that the contract price is the neutral mathematical expectation.
The benchmarks also make different errors. A random walk may miss a sustained shift because it refuses to move from today’s price. Momentum may overreact when a recent shock reverses. Mean reversion may pull too quickly toward a historical pattern during a genuine structural change. When futures beat one of them, the result should be described in relation to that failure mode. The evidence becomes more informative when it says what kind of backward-looking assumption the futures curve improved upon.
The sample determines the question being answered
The Federal Reserve exercise uses monthly data for 12 commodities beginning in 2005. Futures prices are sampled on the last trading day of each month. The comparisons use contracts three, six, and twelve months away from the contract used to measure spot. This design puts the curve into a common forecasting framework while retaining multiple horizons. A three-month comparison asks about a different interval of uncertainty and market adjustment than a year-ahead comparison.
The study groups observations by the ratio of the futures price to the spot price. A ratio close to one represents a curve that is near spot. Ratios materially below or above one represent larger deviations, with separate groups for different magnitudes. This grouping is not a claim that one threshold has a natural meaning in every market. It is a way to ask whether the performance of futures changes when the curve is more visibly separated from spot.
The spot measure and the horizon also deserve attention. The study follows an approach in which the maturing futures contract is used to measure spot. Earlier contracts are then evaluated against that later measure. Readers should not silently substitute a different spot series, delivery convention, or maturity and assume that the published comparison is unchanged. Forecast performance is partly a property of the measurement choices used to define the forecast and its outcome.
The end of the sample is not identical across all comparisons. The table notes report different endpoints by maturity and commodity group: the three-month specifications extend later than the six-month specifications, while the twelve-month energy and metals comparisons end at a different date from the agricultural comparisons. This is a normal consequence of requiring a future observation for each forecast horizon. It means that the observations behind a short-horizon result are not exactly the same set as those behind a year-ahead result.
The design also creates uneven evidence across commodities. A market may rarely produce a large futures-to-spot deviation, so its extreme-deviation cells can contain few or no observations. Industrial metals are specifically described as markets in which meaningful deviations are uncommon. An empty cell does not show that the commodity cannot enter that state; it shows that the specified sample did not provide observations for that comparison. Absence of a populated cell is not evidence of absence of the underlying market condition.
These details are not administrative footnotes. They determine how much weight a reader can place on a pattern. A result that appears across several commodities and maturities is different from a result found in one thin tail of one market. A result that holds for a near-term contract may not automatically carry to a farther-dated contract, particularly when liquidity and physical-market conditions differ. The comparison becomes more credible when its denominator, horizon, population of observations, and deviation state are all visible.
What the design can establish
Within its sample and specifications, the Federal Reserve analysis can establish whether futures had lower relative RMSE than each of the three simple benchmarks in different futures-to-spot states. It can show whether the relationship is stronger when the curve is materially away from spot, whether the pattern varies by commodity or maturity, and whether a benchmark is especially weak in certain states. Those are useful findings because they identify conditions in which the curve may add information beyond recent spot history.
The design cannot, by itself, establish that futures equal the expected future spot price. Relative RMSE is not an unbiasedness test. It cannot show that futures dominate every more sophisticated forecast model, because the comparison is limited to the named benchmarks. It also cannot show that a curve deviation causes future forecast improvement. A deviation may be a marker for an underlying market state—such as supply stress or changing inventory conditions—that affects both the futures quote and the difficulty of forecasting.
Nor should the comparison be read as a universal score for “the futures market.” The study covers 12 commodities, selected maturities, monthly observations, and commodity-specific sample endpoints. It uses simple benchmarks and bins observations by the futures-to-spot ratio. Those choices are appropriate to the question the authors ask, but they bound the answer. Extending it to another commodity, horizon, frequency, or future period requires evidence that the relevant market structure and data-generating conditions are comparable.
There is also a subtle modeling qualification. The mean-reversion coefficients were estimated using the full data sample. The note acknowledges that this gives that benchmark an advantage because the estimated responses reflect actual future price changes; a real-time expanding-window approach could produce a different comparator. This does not invalidate the result. It tells us that the comparison should be interpreted with the stated construction rather than imagined as a perfectly real-time contest among forecasts.
A useful reading habit follows. Whenever a table reports that futures outperform, ask five questions: Compared with which forecast? Over what horizon? For which commodities and observations? How far was futures from spot? And does the result describe relative error, absolute error, or unbiasedness? The answers keep the claim attached to the evidence that supports it. They also prepare us to read the next result, where futures generally improve on several simple methods but not uniformly across every market state.