There is a chart on Atelier Mira’s watchlist this Tuesday: stock cover for Product 41, the oak sideboard that has spent a week selling at 1.8 times its usual pace. The middle of the forecast crosses the reorder point around day 22; the honest window runs from day 16 to somewhere past day 35. The case for showing that window at all is made elsewhere in this library. This piece is about what should sit underneath it: the homework a forecast owes you when you ask it how it knows.
Doubts with names
The width of the band is not an aesthetic choice. It is computed from named sources of doubt, and each one can be interrogated. Demand variance: how much the sideboard’s weekly sales have historically wobbled around their average. Data thinness: the hot streak is one week old, and one week is a thin file; the same signal three weeks in would draw a visibly narrower band. Replenishment risk: Riva Trading’s lead times average ten days and have been slipping, and a stock forecast that ignored the supply side would be half a forecast. Ask why the band is wide and the answer names these, with numbers, in the order they matter. A model that cannot answer that question is not forecasting. It is gesturing.
Eight weeks of history, then the honest part: a forecast that admits how wide it is.
Retrained nightly, graded weekly
The model that drew Tuesday’s band is not the model that drew Monday’s. Every night, as part of the beats, it re-reads what actually landed (the day’s sales, confirmed deliveries, anything returned) and refits. Most nights the band barely moves, which is what you want; a forecast that lurches on every data point is confessing it never knew much. The reckoning is weekly: the reconciliation ledger marks past calls against outcomes, and the misses are written against the model by name. Last week a markdown predicted to move four slow movers moved one, and “too shallow” now sits in the pricing model’s file, where the next markdown proposal has to answer to it. The stock forecast carries its record the same way. A forecast with a documented history of being wrong, and of saying so, is worth more than a fresh one with a perfect record and no memory.
Trust a forecast the way you trust a colleague: not because they have never been wrong, but because you have seen what they do about it.
Two futures, priced honestly
The last section of the homework is the one most forecasts skip entirely: the comparison. A forecast exists to sit under a decision, and Mira’s decision has two branches. The honest move is to price both.
- Act now. Cash commits early to hardwood and fittings, ahead of confirmation that the pace holds. If the hot week fades, Mira carries extra stock of its own bestseller: the cheapest mistake available in the building, but a real one, and the model says so instead of rounding it to zero.
- Wait. Every day of waiting buys evidence and spends option. In the fast branch of the band, cover hits the reorder point around day 16; with ten days of supplier lead time, the order has to be placed by day 6. Today is day zero. Friday is day three. Waiting until Friday costs almost nothing; waiting until the middle of next week quietly closes the safe window in the very branch that made this a question.
Which is how the comparison resolves into something unglamorous: not “order now” and not “relax”, but decide Friday, because Friday is free and Monday is not. The watchlist entry carries that deadline. If the strong week becomes two, Friday’s brief holds a reorder proposal with this chart attached; if it fades, the entry stands down and says so. Either way, Monday’s ledger grades the call.
No naked numbers: a projection never appears without its inputs and their ages. No unexplained width: if the model cannot name what makes the band wide, the width is decoration. No comparison-free advice: a recommendation to act, with the cost of waiting left unpriced, is a preference dressed as analysis.
None of this makes the forecast right. Day 22 may come and go with the sideboards still selling, or the far edge of the band may turn out to be the true story. What the homework changes is the quality of the failure: a forecast that showed its inputs, carried its record and priced both futures fails in a way that teaches, and the nightly retraining makes sure the lesson lands somewhere. The one thing it never asks of anyone at Mira is faith.