Your reported forecast accuracy went up for two quarters and inventory and stockouts both got worse. How is that possible?

Two different quantities are being called accuracy

Nothing in the question is contradictory once you notice that the number in the management pack and the number the warehouse experiences are computed over different things. The reported figure is usually a national or regional total, by product family, by month, compared to actual for that month. Inventory is held for an individual item at an individual location, and it has to cover demand over the time it takes to replenish, which might be a week or six.

Aggregation is a powerful accuracy improver and it improves nothing operationally. Errors across items and locations partly cancel, so a total is genuinely easier to forecast than its parts, and a change that improves the total, such as a better handle on seasonality or a cleaner treatment of trading days, can leave the allocation across items and stores worse than before. The result is exactly what the question describes: the aggregate looks better, and every location is holding stock of the wrong things.

The same effect applies to the horizon. A forecast evaluated one period ahead, at the moment it is most informed, can be excellent while the forecast at the lag that replenishment actually uses is poor. If your supplier lead time is five weeks, the forecast that matters is the one made five weeks before the demand it covers, and it is the only one worth reporting against. Measuring at lag one and buying at lag five is one of the commonest ways an organisation congratulates itself.

So the first move in the interview is to refuse the premise until it is specified. At what grain, over what horizon, against which version of the forecast, and computed how. Three of those four are usually wrong.

The metric quietly chooses the winner

How error is summarised changes which model wins, and the default choice is a poor one for logistics.

Mean absolute percentage error divides each error by the actual, which makes it unstable for small actuals and undefined when the actual is zero. In a distribution network, most item-location combinations are small and many are zero in a given period, so a percentage-based average is dominated by the noisiest, least important cells. A model that hedges towards zero on slow movers scores well and starves the shelves. Weighting instead by volume or value, by dividing the total absolute error by the total actual, gives a figure that reflects the business rather than the long tail, and it is usually the change that reveals the accuracy improvement was never where anyone thought.

Bias is the other half and it is the diagnosis most likely to explain the inventory outcome. Absolute error tells you how far off you were; bias tells you which direction, persistently. A forecast that is wrong by ten per cent in both directions at random and a forecast that is wrong by ten per cent upwards every period have similar absolute error and completely different consequences. The first is buffered by safety stock. The second builds inventory relentlessly on some items and, if the bias is negative, produces sustained stockouts on others, and it will not be visible at all in a report that only shows absolute error. Bias should be tracked by item group, by location and by planner, because a cancelling pair of biases in two regions looks like an unbiased national forecast.

And no accuracy figure means anything without a baseline. The comparison is not to perfection, it is to the simplest thing you could have done: last period's actual, the same period last year adjusted for growth, a moving average. A statistical forecast that cannot beat a seasonal naive baseline on the grain and horizon that matter is not adding value regardless of what its percentage says, and stating that as the acceptance test is what distinguishes somebody who has run a forecasting process from somebody who has evaluated models.

Your history is shipments, not demand

The underlying data problem is that most planning systems train on what was sold or shipped, and what was sold is what you were able to sell. Every stockout truncates the demand signal, and the truncation is systematic rather than random: it happens precisely on the items and weeks where demand was highest.

The forecast therefore learns a demand curve with its peaks shaved off, forecasts lower, drives less stock, produces more stockouts, and shaves the next peak further. It is a slow, self-confirming decline that shows up as improving accuracy, because a forecast trained on censored history predicts censored history well. That is the mechanism most likely to be behind a two-quarter improvement in the reported metric alongside worsening service.

The fix is unglamorous and it is mostly data capture. Record availability at item and location by period, so that a period with a stockout can be excluded or its demand estimated rather than taken at face value. Capture substitutions where you can see them, since a customer who bought the alternative has told you something about both items. Record promotional periods, price changes and any event that moved demand for a reason that will not recur, and hold them as explanatory features rather than letting them contaminate the baseline. And treat new products, which have no history at all, as a separate problem solved by analogue selection and judgement rather than by extrapolation.

Improving the forecast without touching the policy changes little

The other reason accuracy and inventory can move in opposite directions is that the forecast is an input to a policy, and the policy has parameters that were set against the old forecast's behaviour.

Safety stock exists to absorb error, and its size is driven by the variability of forecast error over the replenishment lead time and by the service level you have chosen. Those error parameters are frequently set once, from a historical study, and then left. If the forecast changes, the error distribution changes, and the buffers are now sized for a model that no longer exists. In the benign case you are carrying stock you no longer need. In the case that matches the question, the forecast improved on average and became worse in the tail, and the buffers were never resized to cover the tail that remained.

Reorder review periods, minimum order quantities, pallet and case rounding, and supplier minimums all interact with this too. Rounding a small forecast up to a case quantity is where a great deal of slow-moving inventory comes from, and improving the forecast for those items has almost no effect because the order quantity was never determined by the forecast. Knowing which items are forecast-driven and which are constraint-driven stops you spending a quarter improving something with no leverage.

The override that nobody measures

Finally, the number in the pack is usually the output of a process with human steps in it. A statistical forecast is produced, planners adjust it, sales and marketing add their view of promotions and launches, and a consensus figure emerges from a planning cycle. Each of those steps changes the forecast and only some of them improve it.

The way to find out is to keep every version. Store the statistical forecast, the planner-adjusted forecast and the final consensus forecast separately, each stamped with the date it was made, and then measure all three against actuals at the grain and horizon that matter. This is uncomfortable and it is the single most valuable piece of instrumentation in a planning system, because it converts an argument about judgement into evidence. It usually shows that overrides help substantially on a small number of items where somebody genuinely knows something, such as a confirmed promotion or a lost customer, and hurt on everything else, and that finding lets you narrow the override to where it earns its keep.

It also lets you answer the question in the title honestly. If the statistical forecast improved and the consensus forecast that drove replenishment did not, then the model change worked and the process ate it, which is a different problem from the one anyone was investigating.

Report accuracy at the grain, horizon and version that the buying decision actually uses, split into bias and magnitude, weighted by value, and against a naive baseline. Almost every case of accuracy rising while service falls is visible immediately once the figure is computed that way.