# Forecasts – the Need, the Great Damage, and Using it Right

Forecasting means predicting the future based on data and knowledge gained in the past.

According to the above definition every single decision we make depends on a forecast.  This is definitely true for every managerial decision.

The problem with every prediction is that it is never certain

Treating forecasting as a prophet telling us the future is a huge mistake.  So, we need a forecast that would present what the future might look like, including what is more likely to happen, and what is somewhat less likely, but still possible.

Math taught us that describing any uncertain behavior should have, at the very least, two different descriptors/parameters:  a central value, like ‘the expected value’, and another one that describes ‘the expected deviation from the average’.  This leads the definition of a ‘confidence interval’ where the more likely possible results lie.  Any sound decision has to consider a range of possible results.

While there are several ways to obtain effective forecasts, which could be used for superior decision-making, the real generic problem is the misuse of forecasts.

There are two major mistakes in using forecasts:

1. Using one-number forecasts.
2. Using the wrong forecasting horizon or level of detail.  The generic point is that the exact type of the forecast has to fit the decision that would take the forecast as a critical information input. A similar mistake is using the wrong parameters for computerized forecasts or relying on irrelevant, or poor quality, data.

The use of one-number forecasts

The vast majority of the forecasts used in business display only one number per item/location/period.  There is no indication of the estimated forecasting error.  Thus, if the forecast states that 1,000 units are going to be sold next week, there is no indication whether selling 1,500 is still likely to happen, or only 600.  This distorts the value of the information required for a sound decision, like how much to buy for next week sales.

Any computerized forecast, based on even the simplest mathematical model, includes an estimation of the mean possible deviation from the mean.  Given the expected value of a forecast and turning it into a reasonable range, like plus minus 1.5 or 2 estimated standard deviations or using the mean absolute percentage error (MAPE), yields about 80-90% chance that the actual outcome would fall within that range.

How such a reasonable range is able support decisions?

The two key meaningful information items are the boundaries of the range.  Every alternative choice of a decision should consider both extreme values of the range to calculate/estimate the potential damage. When the actual demand equals the lower side of the range it leads to one outcome, and when the actual demand equals the higher side there is another outcome.  When the demand falls somewhere within the range the outcome also falls between the extreme outcomes. Given both extreme outcomes the choice between the practical alternatives becomes realistic and would lead to better decisions than when no such range of reasonably outcomes is presented to the decision makers.

A simple example:  The forecast states that next week sales of Product X would be somewhere between 1,000 and 1,400 units.  The decision is about the level of stock at the beginning of the week. For simplicity let’s assume that there is no practical way to add X units during the week, or move units to a different location.

There are three reasonable alternatives for the decision:  Holding 1,000 units, 1,400, or going after the mean forecast: 1,200

If only 1,000 units are held and the demand is just 1,000 – the result is perfect.  However, if the demand turns out to be 1,400 there is unsatisfied demand for 400 units.  The real damage depends on the situation:  what the unsatisfied customers might do?  Will they buy similar products, develop a grudge against the company or patiently wait for next week?

When the decision is having 1,400 in stock, the question is:  there might be surplus of unsold units at the end of the week – is it a problem?  If sales continue next week and the units continue to look new, then the only damage is the too early expense of purchasing the 400 units.  There might be, of course, other cases.

What is the rational for storing 1,200 units?  It makes sense only when having a shortage or having a surplus is causing the same damage.  If being short is worse than having surplus, then storing 1,400 is the common-sense decision.  When having surplus is causing the higher damage – let’s decide to store just 1,000.

The example demonstrates the advantage of having a range rather than 1,200 as the one-number forecast, which leaves the decision maker wonder how much the demand might be.

There are two very different ways to forecast the demand.  One is through the use of mathematical forecasting algorithm, based on past results, and performed by a computer.  The other way is using the people closest to the market to express their intuition.  The mathematical algorithm can be used to create the required range, but there is a need to define the parameters defining the range, mainly the probability that the actual will fall within the range.

The other type, where human beings use their intuition to forecast the demand, also lend itself to forecast a range, rather than one number.  Human intuition is definitely not tuned to one number.  But certain rules should be clearly verbalized; otherwise, the human forecasted ranges might be too wide.  The idea behind the reasonable range is that possible, but extreme, results should be left outside the range.  This means the organizational culture is accepting that sometimes, not too often, the actual deviates from the forecasted range. There is no practical way to assess an intuitive 90% confidence interval, as the exact probabilities, even the formula describing the behavior of the uncertainty, are unknown.  Still, it is possible to approximately describe the uncertainty in a way that is superior to simply ignoring it.

We do not expect that all actual results will fall within the range; we expect that 10-20% would lie outside the reasonable range

There could be more variations on the key decision.  When both shortages and surpluses cause considerable damage, maybe Operations should check whether it is possible to expedite a certain amount of units in the middle of the week.  If this is possible then holding 1,000 at the beginning of the week and being ready to expedite 400, or less, during the week makes sense.  It assumes, though, that watching the actual sales in the start of the week will yield a better forecast, meaning a much narrower range.  It also assumes the cost of expediting is less than being short or carrying too much.

Another rule that has to be fully understood is avoiding the use of combined ranges of items/locations for forecasting the demand of product family, a specific market segment, or the total demand.  While the sum of the means is the mean of the combined forecasts, combining the ranges yields a huge exaggeration of the reasonable range.  The mathematical forecasting should re-forecast the mean and the absolute mean deviation based on the past data of the combined demand.  The human forecast should, again, rely on the human intuition.

Remember the objective:  supporting better decisions by exposing the best partial information that is relevant to the decision.  Considering a too wide range, which includes cases that rarely happen, doesn’t support good decisions, unless the rare case might yield catastrophic damage.  Having too wide ranges supports much too-safe decisions, definitely not the required decisions for successful companies.

Warning:  Another related common mistake is assuming that the demand for each item/location is independent of the demand for another item or location.  THIS IS USUALLY WRONG!  There are partial dependencies of demand between items and across locations.  However, the dependencies are not 100%.  The only practical advice is:  forecast what you need.  When you need the forecast of one item – do it just for that item.  When you need the forecast of total sales – do it for the total from scratch.  The one piece of information you might use:  the sum of the mean should be equal to the mean of the sum.  When there is a mismatch between the sum of the means and the mean of the sum, it is time to question the underlining assumptions behind both the details and the global forecasts.

The right forecast for the specific decision

Suppose that a consistent growth in sales raises the issue of a considerable capacity increase, both equipment and manpower.

Is there a need to consider the expected growth in sales of every product?

The additional equipment is required for several product families, so the capacity requirements depend mainly on the growth of total sales, even though some products require more capacity than others.

So, the key parameter is the approximate new level of sales, and calculating back the required increase in capacity.  That increase in sales could also require increase in raw materials, which has to be checked with the suppliers.  There might be even the need for more credit-line to bridge between the timing of the material purchasing, regular operating expenses for maintaining capacity, and the timing of the incoming revenues.

Relying on the accumulation of individual forecasts is problematic.  It is good for calculating the average of the total, but not for assessing the average errors.  Being exposed to reasonable conservative forecast of the total versus the reasonable optimistic one, would highlight the probable risk in the investment and the probable gain.

A decision about how much to store at a specific location has to be based on the individual ranges per SKU/location.  This is a different type of forecast that faces higher level of uncertainty and, thus, should be based on short horizons and fast replenishment and by that deal better with the fluctuations in the demand.  The main assumption of TOC and Lean is that the demand for the next short period is similar to last period, thus fast replenishment according to the actual demand provides quick adjustment to random fluctuations.  Longer term planning needs to consider trends, seasonality and other potential significant changes. This requires forecasts that look further into the future and are able capture the probability for such changes and include them in the reasonable range.

There are also decisions that have to consider the forecast for a specific family of products, or decisions that concern a particular market segment, which is a part of the market the company sells to.

The current practice regarding computerized forecasting is to come up with detailed forecasts for every item and accumulate them based on the need.  The problem, as already mentioned, is that while the accumulation of the averages yields the average of the total, when it comes to ranges the resulting range is a much too wide.

Another practice, usually based on intuitive forecasts, is to forecast the sales of a family of products/locations and then assume a certain distribution within the individual items.  This practice adds considerable noise to the average demand for individual items, without any reference to the likely spread.

Considering the power of today computers, the simple solution is to run several forecasts based on the decision-making requirements

When it comes to human-intuition based forecasts, there is flexibility in matching the forecast to the specific decision.  The significant change is using the reasonable range as the key information for the decision at hand.

Data quality

A special issue for forecasting is to be aware what past data is truly relevant to the decision at hand.  Statistics, as well as forecasting algorithms, have to rely on time-series data from the not-too-close past in order to identify trends, seasonality and other factors that impact future sales.  The potential problem is that the consumption patterns might have gone through a major change in the product, market or the economy, so it is possible that what happened prior to the change is not relevant anymore.

Covid-19 caused a dramatic change to many businesses, like tourism, restaurants, pubs and cinemas.  Other businesses have also been impacted, but in a less dramatic way.   So, special care should be taken to forecast future demand after Covid-19, while relying on the demand during the plague.  The author assumes the future consumption patterns for most products and services will behave differently after Covid-19 relative to 2019. This means the power of the computerized forecasts might go down for a while, as not too much good data will be available.  Even human-intuition forecasts should be used with extra care, as intuition, like computerized forecasting, are slow to adapt to a change and be able to predict its behavior.  Using rational cause-and-effect to re-shape the intuition is the right thing to do.

Conclusions

All organizations have to try their best to predict the future demand, but all managers have to internalize the basic common and expected uncertainty around their predictions and include the assessment of that uncertainty into their decision-making.

Once this recognition is in place, forecasts that yield a reasonable range of outcomes would become the best supportive information, leading to much improved decisions.  At times where the common and expected uncertainty is considerably higher than prior to 2020, organizations that would learn faster to use such range-forecasting will gain a decisive competitive edge.

### Eli Schragenheim

My love for challenges makes my life interesting. I'm concerned when I see organizations ignore uncertainty and I cannot understand people blindly following their leader.

## 2 thoughts on “Forecasts – the Need, the Great Damage, and Using it Right”

1. Ravi says:

can you pls explain “sum of means should equal mean of the sums”?

Like

2. Hi Ravi, the word “mean” is also the synonym for ‘average’. The idea is that if we add all the average durations of the tasks along the critical pass/chain, we’ll get the average duration of the whole chain.

Like