Can a Small Italian Business Forecast Demand Without a Data Science Team?
A practical approach to demand forecasting for Italian distributors and manufacturers, from a simple baseline to machine learning and human planning.
Operations, purchasing and finance leaders
6 min read
2 November 2026
3 sources
Define
Decision + horizon
→Prepare
Stock + demand
→Compare
Simple baseline
→Review
Range + lead time
→Measure
Stock + cost
Test with information available when the decision was made.
A forecast is useful when it improves an ordering decision, not when it produces a convincing-looking chart.
- 01Compare machine learning with a simple seasonal baseline.
- 02Test using only information available at the time of the prediction.
- 03Connect each forecast to a stock or purchasing decision.
Start with the decision, not the algorithm
A food distributor wants to order packaging six weeks before a seasonal promotion. Too little stock interrupts production; too much consumes cash and storage. The useful question is not 'Can AI predict sales?' but 'How many units should we order now, and what happens if demand is higher or lower?'
Istat's 2025 enterprise survey found AI use increased among Italian businesses with at least ten employees, but also identified skills, data quality and cost as barriers for firms considering adoption. A forecasting project should be small enough to be understood and maintained by its owner, not designed around an impressive model name.
Build a baseline that can be beaten
Choose a handful of products with stable identifiers and a consistent planning horizon. Check order dates, stockouts, returns, pack sizes and discontinued lines. A period with no sales because a product was unavailable is not the same as a period with no demand.
Compare the current planner's method with a simple seasonal average or last-year comparison. Only then test a machine-learning model using features genuinely available when the order decision was made. Future revisions, late sales postings and knowledge of the eventual promotion result must not leak into the training or test data.
Turn a range into an action
Show a forecast range alongside lead time, current stock, open orders and supplier minimums. A buyer can then compare the cost of shortage with the cost of excess stock. A predicted increase does not automatically justify ordering if the supplier can replenish quickly or the product is perishable.
Give the planner a way to explain an override: a cancelled customer contract, unusual weather event or known promotion. Keep the original forecast and the final decision so the team can learn whether adjustments improved the outcome.
Prove value over several cycles
Use a chronological backtest, followed by a live shadow period before changing purchasing decisions. Measure forecast error by product group, stockouts, excess stock and planner time. A single average accuracy figure can hide serious misses on the few products that matter most.
Review performance after a supplier change or product launch. If the data is too sparse for a reliable model, a transparent rule with a human owner may be the right first step.
Use this before approving the work.
- 01Name the purchase decision and planning horizon.
- 02Check stockouts and product identifiers before modelling.
- 03Benchmark against a simple seasonal method.
- 04Track business outcomes and planner overrides.
Research and references supporting this article. Interpretation and recommendations are Thalenor’s.
The distributor and planning scenario are constructed; the sources provide adoption and risk-management context rather than proof of forecasting results.
Published and reviewed on 2 November 2026. This article provides general technology and operational information, not legal, tax, medical or regulatory advice.