You’re Not Ordering for Next Week

By Matt Wampler, CEO of ClearCOGS

Quick Answer: A weekly order almost never covers seven days. It has to cover the window from the moment inventory was counted to the arrival of the delivery after the one you are placing, because that is the last moment you can do anything about a shortage. For a typical cycle of counting one night, ordering the next morning, and receiving a few days later, that window is usually somewhere between ten and fourteen days. Sizing the order against seven is the most common reason groups run short before the truck.

Why the number is bigger than it looks

Ask an operator how far ahead they order and most will answer in terms of the cycle. We order weekly. The truck comes Friday.

But the quantity on that order has to survive until the next truck, and the clock did not start when you placed the order. It started when somebody counted.

Walk through a common cycle:

An illustrative weekly ordering cycle. Replace the days with your own.

The count on Tuesday night is the last piece of real information you have. The next chance to correct anything lands the following Friday. That is ten days, and if the delivery slips to Monday or the count happens a day early, it stretches further.

The order is not covering a week. It is covering a week and a half, built on a number that starts going stale the moment it is written down.

What is a forecast horizon?

The forecast horizon is how far into the future a prediction has to reach. In ordering, it is the count-to-next-delivery window described above.

It matters because horizon is the single biggest driver of how hard a forecast is. Predicting tomorrow is close to easy: you know the day of the week, you can see the weather, you know what is on the calendar. Predicting day eleven is a different exercise, and the errors are not comparable.

Most operators size their order with a method built for the easy version of the problem. An average of the last few weeks of usage is a reasonable estimate of a typical day. Stretched across a horizon nearly twice as long as people think it is, it gets thin.

Why longer horizons need more than a bigger number

Here is the part that catches people out, and there is good evidence behind it.

The M4 Competition is one of the largest forecasting experiments ever run. It tested 61 forecasting methods against 100,000 time series, evaluating both point forecasts and the prediction intervals around them (Makridakis, Spiliotis, and Assimakopoulos, International Journal of Forecasting, 2020).

The finding worth an operator’s attention is not about which method won. It is about the intervals. The competition organizers describe most participating methods as considerably underestimating uncertainty, with only the top two of the sixty-one specifying their 95 percent intervals accurately. Serious forecasters, using serious methods, were systematically too confident about how wrong they might be.

If that is true of professional forecasting methods, it is almost certainly true of a spreadsheet that applies a fixed percentage on top of an average. The point estimate may be fine. The allowance for being wrong is the part that tends to be too small, and the longer the horizon, the more that shortfall costs you.

Why a flat percentage buffer is the wrong shape

Most groups add a buffer, and most buffers are a single number applied to everything. Ten percent on top. Or an extra two days.

That is administratively simple and operationally wrong, because items do not vary equally.

A staple that moves steadily every day has tight, predictable usage. A fixed percentage on top of it is mostly waste. An item tied to a promotion, a weekend pattern, or one large customer swings hard, and the same percentage leaves you short. One rule applied to two very different distributions guarantees you are over-buffered on half your list and under-buffered on the other half.

The better shape is a buffer sized to each item’s own variability across the actual horizon. In practice that means asking, for this item over these eleven days, what is a realistic high end, not what is the average plus ten percent.

Operators already do a version of this instinctively. They add more cushion to the thing they cannot run out of and less to the thing they can substitute. Making it explicit, and per item, is what turns instinct into something that survives a change of staff.

What to do about it

Four steps, none of which require new software to begin.

1. Write down your actual horizon. Count day to delivery day to next delivery day. Most groups discover the real number is three to five days longer than the one they have been using.
2. Check whether your buffer covers that window or the shorter one. If your buffer was set as a percentage years ago and nobody has revisited the delivery schedule since, these have probably drifted apart.
3. Split your items by volatility, not by cost. Take the twenty items that most often run short and look at how much their weekly usage actually moves. That spread, not a flat percentage, is what should set their buffer.
4. Shorten the horizon where it is cheap to do so. Counting closer to the order day, or adding a second delivery for a handful of high-volatility items, reduces the forecasting problem rather than trying to out-predict it. A shorter window is worth more than a better model.

That last point is the one worth sitting with. Everybody reaches for a more accurate forecast first. Frequently the cheaper win is a smaller gap between the last good information and the decision.

Frequently Asked Questions

How many days of inventory should a restaurant order?

Enough to cover the period from your inventory count to the delivery after the one you are placing, plus a buffer sized to how much each item’s usage varies. For a weekly cycle that typically lands between ten and fourteen days, but you should calculate it from your own count, order, and delivery days rather than using a rule of thumb.

Why do we run out before the next truck even though we ordered a full week?

Usually because the order was sized for seven days when the real window was longer. The gap between the count and the order, and between the order and delivery, both add days that often go uncounted.

Should the buffer be the same for every item?

No. Buffers should reflect how much each item’s usage varies. Steady items need very little. Items tied to promotions, weather, or a small number of large orders need considerably more.

Is it better to order more often or forecast better?

Ordering more often is usually the cheaper improvement, because it shortens the horizon and every forecast gets easier as the horizon shrinks. It is not always practical given delivery minimums and fees, but it should be evaluated before assuming the answer is a better model.

Does counting inventory more often fix this?

It helps, but only by moving the start of the window closer to the order. Counting twice as often without changing the delivery schedule shortens the horizon a little and adds a lot of labor. Changing when you count relative to when you order is usually the higher-value change.

The order you place on Wednesday morning is not a bet on next week. It is a bet on a window that started before you placed it and ends after the next truck, and for most groups that window is meaningfully longer than the one their math assumes.

Measure the real horizon first. Then decide whether the answer is a better forecast, a bigger buffer on the right items, or simply a shorter gap between the count and the decision.

Let’s Talk

Sources

• Makridakis, Spyros, Spiliotis, Evangelos, and Assimakopoulos, Vassilios. The M4 Competition: 100,000 Time Series and 61 Forecasting Methods. International Journal of Forecasting, 36(1), 54–74. 2020. sciencedirect.com