The most common quantitative error in commodities is regressing two price series on each other in levels and reading the r-squared as evidence of a relationship. It is worth understanding exactly why this fails, because the failure is severe, systematic, and produces very convincing numbers.
Prices, inventory levels and the dollar index are all close to integrated of order one: their changes are roughly stationary but their levels wander without reverting. Two independent series with that property, regressed on each other, will produce a high r-squared and a significant t-statistic a large fraction of the time. This is the Granger-Newbold spurious regression result, and it is not a small-sample curiosity.
So the first rule is to difference both legs. Regress changes on changes, not levels on levels. A relationship that survives differencing is a relationship between the movements, which is what you almost always wanted anyway. A Durbin-Watson statistic on the residuals is a useful additional gate: strong positive serial correlation in the residuals is the signature of a levels regression pretending to be a relationship.
The second rule is subtler and cost petral a working feature. A fit-quality threshold answers the question 'is this real', not 'is this strong', and the answer depends on the sample size. A fixed r-squared floor conflates the two. A floor of 0.25, calibrated for levels and never re-derived after the fit moved to changes, rejected four of five genuinely live relationships, including crude against the dollar at an r-squared of 0.074 on 299 observations, which is roughly twenty times what unrelated series produce at that length.