The logistics firm that was very good at predicting last week

The situation

Last-mile logistics in the UAE runs on a short planning horizon and a long tail of peaks: retail promotions, national holidays, Ramadan, and the e-commerce events that now rival all of them. Getting the fortnight ahead roughly right determines whether an operator staffs from its own pool or pays agency rates.

The operator had understood this and invested accordingly, buying a forecasting capability two years before we arrived.

The problem

The forecast was consistently wrong in a particular way. It tracked ordinary weeks well and under-predicted peaks, which is the expensive direction: an under-forecast peak is covered with agency drivers booked at short notice.

Two years of effort had gone into the model. Features had been added, algorithms swapped, a consultancy engaged to tune it. Accuracy had not moved.

Meanwhile the planning team maintained a spreadsheet. Everyone knew about it. The spreadsheet consistently beat the model, and management had spent eighteen months trying to get planners to stop using it, on the reasonable grounds that an organisation should not run on one analyst's workbook.

What we did

We began by asking to see the spreadsheet rather than the model, which was not what the brief asked for. The brief asked us to improve forecasting accuracy. Our position was that a manual process outperforming a funded model for two years is the most informative thing in the building, and that nobody had treated it as evidence.

The senior planner walked us through it. Her sheet ignored the system's consignment volume field entirely and used a count she derived from manifests. Asked why, she said the system field had "gone wrong at some point" and she had stopped trusting it.

She was right. A data migration two years earlier, when the operator changed transport management systems, had mapped multi-parcel consignments in a way that collapsed them to a single unit. Roughly a fifth of records were affected, and the affected proportion rose during peaks, because peaks are when multi-parcel consignments are most common.

The model was not wrong. It was being trained on a series that systematically understated exactly the periods it was expected to predict, and every tuning effort had been an attempt to fix downstream what was broken upstream.

This is an uncomfortable finding to deliver. It meant two years of model work had been unnecessary, and the person who had spotted it first had been repeatedly asked to stop.

We corrected the historical series, rebuilt the affected records from manifest data, and put validation in place at ingestion so a future migration cannot silently change the meaning of a field. Only then did we retrain, and the model needed almost no adjustment.

The outcome

Mean absolute error on the fourteen-day volume forecast fell from about thirty-one per cent to about nine.

Agency driver spend dropped roughly twelve per cent in the first peak after the correction, which is the number the operations director cared about and the only one that appeared in the original business case.

The planner's spreadsheet was retired, not because she was told to, but because the system finally agreed with it.

The broader lesson the operator took was about where to look. A model is a lens. Polishing a lens that is pointed at the wrong thing produces a sharper picture of the wrong thing, and for two years everyone had been polishing.