Store labour forecasting, then and now
We built store-level demand and labour forecasts for automotive-service retailers. Foundation models and new research on scheduling and churn change what we would build today.
From 2019 to 2023 we built forecasting models for national automotive-service and tire retailers: demand by store and service type, and the labour needed to meet it. The deliverable was production-ready Python, a model report with agreed KPIs, and a knowledge-transfer session so the client's team could run and retrain it.
The models worked. Planners started from a forecast instead of last year's spreadsheet. But looking back, three things took most of the effort, and all three look different in 2026.
1. One model per series is no longer the default
We trained and tuned models per segment, with careful feature engineering for promotions, holidays, weather and seasonality. Much of the project time went into maintaining that zoo.
Time-series foundation models change the starting point. Amazon's Chronos-2, released last October, does zero-shot univariate, multivariate and covariate-informed forecasting, using known future inputs such as promotions and holidays. It ranks first among pretrained models on the GIFT-Eval benchmark, and Chronos models have been downloaded more than 600 million times.
Today we would start with a foundation model as the baseline and fine-tune or add custom models only where they beat it on held-out stores.
2. The forecast is not the goal; the schedule is
We delivered forecasts. What managers actually needed was a schedule people could live with. New research makes that explicit. In Harvard Business Review, Santiago Gallino and Borja Apaolaza analysed 280 million shifts worked by 1.3 million employees at 20 U.S. retail chains. Retailers that gave two to three weeks of schedule notice averaged about 5% monthly attrition, versus 7% to 8% for those giving less than a week. Using LASSO regression, they narrowed 166 scheduling variables to the few that predicted turnover, and the drivers differed by site.
A forecast that is accurate but arrives too late to post a stable schedule is costing more than it saves.
3. The last mile was manual
Turning forecasts into staffing decisions was spreadsheet work: export, adjust for local knowledge, copy into the scheduling system. Today an agent can do the routine part: pull the forecast, apply the store's rules, draft a schedule with the required notice, flag the stores where the forecast moved sharply, and explain the changes in plain language to the manager who approves it.
What we would build now
- A foundation-model baseline, with custom models only where they earn their keep.
- Forecasts produced on the schedule-notice horizon, not just next week.
- A scheduling agent that drafts, explains and flags, with the manager approving.
- Turnover tracked as an outcome metric next to forecast accuracy.
Same problem, much less code, and closer to what the business actually needs.
Sources
- Introducing Chronos-2: From univariate to universal forecasting, Amazon Science, October 20, 2025.
- The Solution to Service-Worker Churn, Santiago Gallino and Borja Apaolaza, Harvard Business Review, March 2026 (syndicated copy).