The forecast gate that got worse after seeing raw errors

A retailer knows next week includes a promotion and a holiday, but has almost no past week with that exact combination. One option is to ignore history. Another is to retrieve a promotion week and a holiday week separately, then decide whether either deserves to move a base forecast. FARE takes the second route.

Gu and colleagues make that decision with two small logistic gates. The striking result is not merely that the gates help. It is that seemingly better reliability inputs can hurt: on all four datasets, raw series-level errors and pooled global errors are worse than omitting the error history entirely. The useful signal appears only after sparse series estimates are shrunk toward a global mean.

A known future factor is information available before the forecast, such as a promotion, holiday, discount, or SNAP benefit day. FARE searches historical 56-day sales windows twice: one ranking rewards similarity plus agreement on factor one, and the other does the same for factor two. Alternating between rankings reserves four of eight reference slots for each factor and removes duplicates.

That matters when the conjunction is rare. Requiring promotion and holiday can return no precedent; a single joint score can also spend all eight slots on smooth-looking histories that mismatch one factor. The paper reports factor-wise Top-8 beating joint Top-8 on every dataset under the same retrieval budget.

The candidates are a backbone forecast B, a shape-only reference S, and a factor-aware reference F. The first sigmoid gate α asks, “what should the historical reference be?” It forms R = S + α(F − S). The second gate β asks, “how far should we leave the backbone?” It forms P = B + β(R − B), clipped at zero.

def shrunk_error(error_sum, count, global_mean, lam=14):
    return min(10.0, (error_sum + lam * global_mean) / (count + lam))

def fuse(backbone, shape, factor, alpha, beta):
    reference = shape + alpha * (factor - shape)
    return max(0.0, backbone + beta * (reference - backbone))

def weights(alpha, beta):
    return (1 - beta, beta * (1 - alpha), beta * alpha)

In the hand-verified example B = 100, S = 80, F = 120, α = 0.75, and β = 0.40. The final forecast is 104. Its effective weights are 0.60 on B, 0.10 on S, and 0.30 on F. On the illustrative three-day demand vector, pooled WAPE is 5.00%.

Candidate reliability is the normalized absolute error accumulated from completed development weeks. A series with n observations and error sum E receives estimate (E + λμ)/(n + λ), where μ is the previously eligible global mean and λ = 14—two weeks of seven daily forecasts. With no local history the estimate is μ; as n grows it moves toward the local mean.

At seven observations with raw mean 2.0, shrinkage gives 1.0 rather than trusting either 2.0 or 0.5. This is empirical-Bayes logic in a small formula: borrow strength when a series is thin, then surrender that borrowing as evidence accumulates.

The paper does not show that two factors are universally enough. Its evidence covers seven-day horizons and exactly two known factors on four retail datasets. Favorita deliberately removes some complete promotion–holiday combinations from the reference library, but the method is not tested when every constituent factor pattern is absent.

The gates and their error statistics are frozen during test, while the retrieval library can grow as completed test weeks become available. Search cost also grows with that library: the tiny 26–32 parameter count does not include retrieval. Most importantly, if B, S, and F all miss the same demand shock, no convex blend can recover it.

A clean reproduction should first implement legal time boundaries, factor-wise retrieval, the λ = 14 estimator, and the two gates around a simple seasonal or LightGBM backbone. The crucial diagnostics are support-stratified reliability curves and an ablation that compares no history, global history, raw local history, and shrinkage—not merely the final WAPE.

Next, vary λ by support, let the estimator adapt under drift, and test more than two factors without exploding the retrieval budget. For a foundation-model baseline, the architecture connects naturally to the site's Transformer walkthrough; the important lesson lives outside the backbone: scarce performance records require regularization just as much as scarce training labels do.