Why better forecasts alone do not create better supply chains
Forecast accuracy is an input. The value shows up in the decisions made with it: what to buy, make, hold and promise.
A familiar pattern in supply chain improvement goes like this. A forecasting project improves accuracy on the dashboard. A year later, inventory, service levels and expedite spend look much as they did. The project gets labeled a disappointment, and the next one promises a better model.
The model is rarely the whole problem. A forecast does not move stock. A buyer, planner or scheduler does, using rules that turn a forecast into a purchase order, a production plan or a promise date. If those rules ignore uncertainty, if the forecast leans consistently one way, or if people override it without saying why, accuracy gains never reach the warehouse or the profit and loss statement.
The decision is the unit of value
Follow a forecast through the chain. It becomes a net requirement once stock and open orders are netted off. The requirement becomes an order quantity once minimums, pack sizes and lead times are applied. The order becomes stock, and the stock becomes either a filled order or days of cover sitting in a warehouse. Each step has its own rule and its own owner, and each one can waste the work done upstream.
Forecasts also differ in how much they matter. A one-point accuracy gain on a made-to-order item with a long customer lead time changes very little. The same gain on a high-volume item with a twelve-week import lead time and a short shelf life can change a great deal. A single company-wide accuracy figure hides this. The useful question is narrower: for this decision, on these items, did a better forecast produce a better result?
Forecast at the level and lag the decision uses
Every decision consumes the forecast at a particular level of detail and a particular distance ahead. A purchase order for an imported item uses the forecast by item, made one lead time before the goods arrive. A capacity booking uses the forecast by product family, months ahead. An allocation uses the forecast by item and location for the next few weeks.
Accuracy reports often ignore this. They measure last week's forecast for this week, at a total level, and declare it good. For an item with a twelve-week lead time, the forecast that drove the order is the one made twelve weeks earlier. Last week's number arrived too late to change anything. Measure accuracy and bias at the level and lag each decision actually uses, and the picture often looks quite different.
Bias does more damage than error
Forecast error measures how far off the forecast was, ignoring direction. Bias measures whether it leans one way. Two forecasts can have identical error and very different consequences.
Bias usually has a source that a better model will not fix. Sales targets typed in as forecasts push it up. Fear of excess stock pushes it down. Promotions and price changes that never reach the planning team show up as unexplained misses. The fix is to measure bias by product group and by who last touched the number, then remove the cause.
Plan on ranges
A single-number forecast hides what a planner most needs to know: how wrong it could be. A range makes that visible. A common way to express it uses three levels. The P50 is the level demand is as likely to fall below as above. The P90 is the level demand stays below nine times out of ten. The P10 is the level it stays below only one time in ten.
Different decisions should plan to different points in that range. Choosing the point is a business call about the cost of running short against the cost of holding too much.
- A critical component that stops a line when it runs out may justify planning near the P90.
- A short-shelf-life item, where excess turns into write-offs, may be planned nearer the P50.
- A capacity booking can be placed at the P50 with an agreed option to flex toward the P90.
Write these choices down by product class, and review them when costs change. The width of the range matters too. When it widens, the sensible response is often a smaller, reversible commitment: buy in smaller lots, delay the final decision, or keep a second supplier warm.
Overrides need reasons
Planners and sales teams know things a model cannot see: a lost tender, a new store listing, a customer's plant shutdown. Overrides that carry that knowledge are valuable. Overrides made from habit, or to hit a target, are costly. Without a record, nobody can tell the two apart.
A simple discipline fixes this. Every override records the original value, the new value, a reason from a short list plus free text, who made it and when. After the period closes, compare each override with what actually happened. This practice is often called forecast value added. It asks whether each step in the process, including the statistical model and each human adjustment, moved the forecast closer to actual demand or further away.
Run the comparison by reason code, and keep it free of blame. Over a few cycles, patterns appear. Overrides citing new listings may add value. Overrides citing "gut feel" on stable items may not. The goal is to keep the overrides that help and stop spending planner time on the rest.
Measure the decision alongside the forecast
Keep measuring forecast accuracy and bias. Then add a second layer of measurement at the level of the decision. For each decision type that uses the forecast, record what was decided, on which forecast and range, what happened and what it cost.
- Purchase quantities: stock-outs, days of cover and excess stock by product class.
- Production plans: schedule changes inside the frozen window, and the overtime or changeovers they caused.
- Allocation in short supply: fill rate for priority customers and penalties paid.
- Promise dates: the share of promises kept.
This shows where a forecast improvement would pay, and where the real bottleneck is a rule, a stale lead time or a minimum order quantity that no forecast can fix.
Five questions for your next S&OP review
- What is our forecast bias over the last 13 weeks, by product class, and where does it come from?
- Which decisions use the forecast, and which point in the range should each one plan to?
- How much of the forecast was overridden last cycle, and did the overrides make it better or worse?
- Do our order quantities change when uncertainty changes, or only when the single-number forecast moves?
- Which decision cost us the most last quarter, and was the forecast really the cause?
In FLX4's loop of Predict → Decide → Execute → Learn, the forecast belongs to Predict, and the value is made in Decide. PlanningNest shows each forecast as a range (P10, P50 and P90) and records every planner override with the original value, the new value and the reason.