You forecast every store and also the national total, and the store forecasts do not sum to it. What do you do?

Coherence is a requirement, not a tidiness concern

The moment the same quantity is forecast at two levels, someone downstream will compare them. Finance builds a revenue plan from the national number, the supply chain allocates stock from the store numbers, and if the two disagree by three per cent then one of them is wrong and nobody knows which. The organisation resolves this politically rather than statistically: each team uses the version that suits it, the forecast stops being a shared input, and the model's accuracy becomes irrelevant to the argument.

So the deliverable is not a set of accurate forecasts. It is a set of forecasts that add up and are as accurate as possible subject to adding up. State it that way and the rest of the answer follows, because reconciliation stops looking like a post-processing step and becomes the design.

Where the signal lives

The reason this is not solved by simply forecasting one level is that the levels have genuinely different statistical properties.

A single store in a single category is a thin series. Weekly units might be in the single digits, with many zeros, and a large share of the variation is irreducibly random — a few customers behaving individually. Seasonality is present but noisy, and a promotional effect is hard to estimate from so few observations.

The national total for the same category is a sum over hundreds of such series, and the independent noise largely cancels while the common structure survives. The trend is visible, the annual seasonality is clean, and price elasticity is estimable. It is a much easier series to forecast well and the accuracy figures will be much better.

That asymmetry is the whole tension. The detail is where the decisions are made and the signal is at the top.

flowchart TD
  N[National total<br/>clean trend and seasonality] --> R1[Region North]
  N --> R2[Region South]
  R1 --> S1[Store 101<br/>noisy weekly units]
  R1 --> S2[Store 102<br/>noisy weekly units]
  R2 --> S3[Store 201<br/>intermittent demand]
  R2 --> S4[Store 202<br/>noisy weekly units]

Read it in both directions: aggregating upwards cancels noise, and allocating downwards has to invent the split that aggregation destroyed.

Bottom-up, top-down, middle-out

Bottom-up forecasts each leaf and sums. It is coherent by construction, it preserves whatever local knowledge each series carries, and a store that genuinely behaves differently is represented. The cost is that you are forecasting hundreds of noisy series and adding their errors together, so the national number can be materially worse than a direct model of the national series would have been. There is partial cancellation if the errors are independent, but they rarely are — a shared seasonality misspecification biases every leaf in the same direction and the bias adds up rather than cancelling.

Top-down forecasts the aggregate and splits it by proportions. The top number is as good as your best model can make it, and the leaf numbers inherit its stability. The problem is the proportions. Historical average share is the usual choice and it is a forecast in disguise: it assumes each store's share of the total is stable, which is false whenever stores are opening, refitting, changing catchment or being cannibalised by a new site down the road. Top-down also systematically flattens genuine local behaviour, so a store with a different seasonal shape gets the national shape scaled down, and the planner who knows better stops trusting the system.

Middle-out picks an intermediate level where the series are still thick enough to model well, forecasts there, then aggregates upwards and disaggregates downwards. It is frequently the pragmatic answer, and the level to choose is the lowest one at which the series are reliably forecastable rather than the level the org chart suggests.

Optimal reconciliation

The modern framing removes the choice. Forecast every series at every level independently and as well as you can — these are called base forecasts and they will not be coherent. Then apply a linear projection that maps the full vector of base forecasts onto the nearest set that satisfies the aggregation constraints, weighting the adjustment by how reliable each base forecast is.

The appeal is that information flows in both directions rather than one. The clean national seasonality informs the store forecasts, and a store with a genuinely divergent trend pulls the total. Bottom-up and top-down are both special cases of this projection with particular weights, which is a useful thing to be able to say, and the general version can be shown to be no worse than either under its assumptions. The weighting is the part that matters in practice, and it is estimated from the in-sample errors of the base forecasts, with structural approximations used when the full error covariance cannot be estimated reliably at scale.

The operational costs are real. You now maintain a model for every node rather than one level, the reconciliation step needs the hierarchy encoded as a matrix and kept in step with a store estate that changes monthly, and explaining to a planner why their store forecast moved when no store data changed is a genuine change-management problem. That last one is worth raising unprompted, because it is the reason reconciled forecasts get overridden in practice.

Which level you are judged on

The instinct is to report accuracy at the leaf level, because that is the finest detail. It is usually the wrong emphasis, and being clear about why is what separates a strong answer here.

Leaf-level error is dominated by irreducible randomness. A store selling four units a week will show a large percentage error no matter how good the model is, and improving the model barely moves that number, so it is a poor signal about whether your work helped. Percentage-based metrics are also unusable there, because a week with zero actual demand has no defined percentage error at all.

Meanwhile the totals drive the decisions with the largest financial consequence: how much stock enters the country, what the revenue plan says, how much manufacturing capacity is booked, what the cash requirement is. An error at the top is not offset by anything, whereas an error at one store is partly absorbed by the opposite error at the next store and by whatever safety stock the network holds.

So report at every level and be explicit about which level each consumer is served by. Use absolute or scaled error measures rather than percentages at the leaves, and quote the aggregate error separately rather than letting it be implied by an average over hundreds of series — an average leaf error can improve while the total gets worse, and the total is what someone has committed money against.

A forecast the organisation cannot reconcile is a forecast the organisation will not use. Decide which level each consumer reads, make the numbers coherent by construction, and report accuracy at the aggregate separately because that is where the money is committed.