Forecast Accuracy Is Declining: What to Check First

When forecast accuracy drops, the first suggestion is almost always a new model or a new tool. It is rarely the right one. In most businesses a decline in accuracy is caused by a change in data, process or behaviour - and a new model applied to the same inputs will reproduce the same decline within two cycles.

Work through the causes in order of likelihood and cost to fix. The list below is roughly that order.

1. Confirm the measurement before believing the decline

A surprising share of "accuracy problems" are measurement artefacts. Check, in this order:

• Aggregation level. Accuracy at item-location is always worse than at category-region. If someone changed the reporting level, the metric moved without the forecast changing at all.
• Lag. A one-month-ahead forecast and a three-month-ahead forecast are different measurements. Compare like with like.
• Formula. MAPE, weighted MAPE and forecast-value-added answer different questions and react differently to low-volume items. MAPE on intermittent demand is close to meaningless.
• Population. New items, promotional periods and discontinued items entering or leaving the calculation base can shift the number several points on their own.

If the definition changed, stop here and restate the history on the new definition before drawing any conclusions.

2. Distinguish bias from error

Error is how far off you are. Bias is whether you are consistently off in one direction. They have different causes and different fixes.

High error with near-zero bias usually means genuine volatility - the response is buffering, segmentation and postponement, not modelling. Persistent bias in one direction is a process or behaviour issue: sales targets flowing into the demand plan, planners hedging against stockouts, or overrides being applied to protect service. Bias is the more damaging of the two because it accumulates as inventory or lost sales.

Error costs you buffer. Bias costs you money in one direction, every cycle, until someone names it.

3. Inspect the demand history

Statistical forecasts inherit whatever is in the history. Check whether:

• One-off events - a large tender, a stock-build ahead of a price rise, a competitor outage - remain in the baseline unflagged;
• Stockouts have been recorded as low demand, so the history reflects what you could ship rather than what customers wanted;
• Promotions are separated from base demand, or blended into it;
• Product hierarchy changes - renumbering, packaging changes, item consolidation - have broken the continuity of the series.

Cleansing history is unglamorous and usually delivers more accuracy improvement than a change of algorithm.

4. Look at overrides and their outcomes

Manual overrides are legitimate - a planner often knows something the model cannot. The problem is unmeasured overrides. Log every override with an owner and a reason, then compare the overridden forecast against the statistical baseline over time. Forecast value added answers the question directly: is human intervention making this better or worse? In many organisations a meaningful share of overrides subtract accuracy, and nobody knows because the comparison was never run.

5. Check what changed in the business, not the model

Pricing changes, a new channel, a large customer switching order patterns, a service change, a new competitor, a change of route to market. If accuracy declined sharply in one month for one segment, the explanation is almost always a business event, and the fix is an input to the plan rather than a statistical adjustment.

6. Only then consider the model

If measurement is sound, bias is controlled, history is clean, overrides are adding value, and no business event explains the shift - then look at segmentation of forecasting methods. Different demand patterns need different approaches: smooth items respond to exponential smoothing families, intermittent demand needs intermittent-specific methods, and new or short-life items need attribute-based or judgement-led approaches. A single model applied to a mixed portfolio is a common and avoidable cause of mediocre accuracy.

Make the fix stick with cadence

Accuracy is maintained by a routine, not a project. In the demand review, report accuracy and bias side by side, by segment, with named owners for the largest deviations. Review the top ten misses by value, not by percentage error - this keeps attention on the items that actually cost money. Track forecast value added so that the discussion is about whether the process improved the number, rather than whether the number was good.

Frequently asked questions

What accuracy should we expect?

It depends entirely on aggregation level, lag and demand variability, which is why cross-company comparisons mislead. Track your own trend, by segment, on a stable definition.

Is one metric enough?

No. Pair an error metric with a bias metric. Error alone hides the systematic over- or under-forecasting that drives inventory and service problems.

Will a better tool fix this?

Only if the constraint is genuinely computational. If the constraint is dirty history, unmanaged overrides or target-driven bias, a new tool inherits all three.