Forecasting fundamentals

Factories, buyers and trucks all need time to react, so almost every supply chain decision is made against a forecast rather than real orders. A forecast is a best estimate of future demand, and every forecast is wrong. The useful questions are how wrong, and in which direction.

Most demand histories are a mix of a few patterns:

• Level: the typical amount sold right now.
• Trend: steady growth or decline, like a brand gaining shelf space.
• Seasonality: a pattern that repeats on the calendar, such as ice cream in summer, sweets before Diwali, or the weekend peak in a café.
• Noise: the leftover randomness that no model will explain.

Before choosing a method, start from the decision the forecast feeds, because that sets the horizon and the level of detail:

• Replenishing a warehouse needs weekly numbers per item and location, a few weeks to a few months ahead.
• Planning plant capacity needs monthly numbers per product family, several quarters ahead.
• Budgeting needs yearly numbers per category.

Forecasting at the wrong level gives a precise answer to the wrong question. A useful habit is to build the simplest possible forecast first, such as “next week = last week” (the naive forecast). Any method you adopt later has to beat it.

People also separate forecasting (a statistical estimate from data) from demand planning (the business process that adds market knowledge, promotions and judgment, then agrees one demand plan for supply to work from).

A repeatable process

1. Define the decision. What will the forecast drive? How far ahead must it look (at least the time needed to act on it), at what level of detail, and how often will it be updated?
2. Get the right data. Forecast demand, not shipments: when stock runs out, shipments understate what customers wanted. Flag promotions, stockouts and one-off events so they don’t distort the pattern.
3. Look before you model. Plot the history. Is there a trend? A seasonal pattern? A sudden break, such as a lost customer?
4. Set a baseline. Start with the naive forecast (next period = last period) or the seasonal naive (same period last year).
5. Fit candidate methods and test them on data they have not seen.
6. Forecast, monitor and review. Track error and bias every cycle, and investigate when they drift.

Testing on unseen data

The honest test is a backtest: pretend it is week 3, forecast week 4, compare with what happened, move on a week and repeat. Never judge a method on the data it was fitted to; almost anything looks good on its own training data.

Horizon, lag and level

• Horizon must cover the time it takes to act. If a supplier needs eight weeks, the forecast for week nine onward is what drives the order.
• Lag is how far ahead a forecast was made. Measure accuracy at the lag the decision uses: the forecast made eight weeks before, not the one made last Monday.
• Level should match the decision. Forecast where the signal is strongest, then add up or split down carefully; the hierarchical forecasting stop shows how.

A forecast is not a target or a plan

Three numbers often get mixed up:

• Forecast: what we expect to happen.
• Target or budget: what we want to happen.
• Plan: what we will do (produce, buy, ship), given the forecast and our constraints.

When a forecast is quietly set equal to the sales target, it becomes biased by design, and the supply chain builds stock for sales that were only hoped for. Keeping the three separate, and showing the gap between them, is one of the most valuable habits in planning.

Why aggregates are easier

Forecast error shrinks relative to volume when you add things up. A product family is easier to forecast than each of its SKUs, a month easier than each of its days, and a country easier than each store, because independent errors partly cancel (the same square-root effect behind risk pooling). That is why capacity is planned on families and months while replenishment works with SKUs and weeks, and why the detailed levels need safety stock.

Error grows with horizon

The further ahead you forecast, the more can change, so errors grow with the horizon. The practical consequences: commit as late as you can, keep long-range plans at aggregate level, and shorten lead times, because a shorter lead time means a nearer, more accurate forecast is enough.

The main families of methods

Large forecasting competitions have repeatedly found that combining several methods, even by simple averaging, is more accurate on average than picking a single best one, because their errors partly cancel.

Data traps

• Censored demand: sales capped by stockouts understate demand.
• Shipments instead of demand: customer orders swing with their own stocking decisions.
• Breaks: a pandemic, a lost customer or a channel shift makes old history misleading. Shorten the history or model the break.