Košík Demand Forecasting

Context

At Košík, logistics and spoilage dominate margin. Demand forecasting drives how much to stock, pick, and route - daily decisions with millions of euros on the line.

Problem

One analyst owned demand prediction in a maze of spreadsheets. The process was brittle, person-dependent, and hard to improve. Operations needed forecasts they could run every day without a single irreplaceable expert.

My role

As analyst (2019-2021), I owned replacing spreadsheet planning with a practical Python forecasting pipeline and getting it into daily production use.

Constraints

• Trust: Ops would not switch overnight; the model had to run alongside human judgment first.
• Messy reality: Seasonality at multiple horizons, weather, holidays, and operational quirks in the history.
• Scale: Small forecast errors compound into waste and missed sales at national volume.

What I did

• Consolidated historical demand data across horizons (next day, few days, week ahead).
• Built a custom Python model with annual, monthly, and weekly seasonality, weather inputs, and operational patterns in the data.
• Ran parallel forecasts with the incumbent expert; tuned until the combined approach beat the old setup by about 5-10%.
• Piloted, then productionised so the team could use it daily.

Key decisions

• Side-by-side proof: Validate against the spreadsheet mastermind before cutting over.
• Practical model over perfect ML: Ship something ops could trust; leave advanced ML to a later dedicated team (which happened).

Result

Forecasting moved from one person's spreadsheets to a repeatable system in production. Operating costs improved on the order of ~5% at the time; the ML team later took the pipeline further. I learned how data work earns its keep only when operations actually runs it.

What this shows

This project is a good example of the kind of work I do best: taking messy operational planning (spreadsheets, seasonality, warehouse reality) and turning it into a productized system teams can actually run every day.