The following is the second in a two-part series covering the art and science of  demand planning. Read the first part on the art of planning here.

Why do we Need to Forecast?

Companies need to understand what the future will bring to make the right decisions: What is my optimal footprint? What is the right price based on future expected volume? How much capacity is needed, and which inventory investments will yield the best returns? Using a statistical forecast provides a good starting point to answer all these questions.

However, not everyone has the background to understand the math behind the process, the underlying assumptions used in the forecast, or the best way to integrate this information into demand planning decision making. This is why it is important to simplify the statistical forecasting principles and assumptions at a high level so everyone actively participate in the discussion and help build the demand plan.

The Science Behind Statistical Forecasting

“The only way to predict the future is to understand the present.” – Isaac Asimov

The first question that needs to be answered is: What are you trying to predict? Forecasting order entry or sales will provide different outcomes and maybe both are needed, but for different reasons. My experience is that to drive the supply chain process you are better off using order entry as an input, since sales includes variability related to supply availability. But if there is a conscious thought process behind the decision of what is being used, it should not be a problem. Let’s review a quick example to clarify this situation.

A few years ago, during a demand discussion among colleagues, we were seeing a high level of volatility in a certain SKU. The Demand Planner was not able to explain it well as the item had steady sales in the past. Then the Supply Planning Manager stepped in and mentioned that we had a recent shortage in this product and the peak in sales should be due to a big lot coming in and being shipped out in the same month.

These circumstances lead us to review the process and the data we were using, as we were trying to determine what we could sell if there was not any supply constraint. The best set of data to do this was orders entered with the requested date from the customer. Once we used these figures, the pattern was much smoother as order entry had the information of what the customer needed, and the sales information had the information of when we were able to provide it. These subtle differences in the input of the process provide a very different output.

Now we can move to discuss the process. The intention of this section is to understand from a ten thousand foot view the components of the models and the high-level assumptions behind them. There will be complex calculations involved and we will not get into the details behind the math. After all, the experts can handle this very easily, but it is the understanding of these concepts in general that helps us to drive constructive discussion and decision-making.

There are a lot of different methods that try to predict the future including a simple moving average, exponential smoothing, econometric models, linear programming methods or machine learning algorithms. We will focus on time series – a series of sequential data points ordered by time. The two main methods we will review are exponential smoothing and linear regression because they are the easiest models to understand and the most widely used.

Exponential Smoothing

The first technique is called exponential smoothing. This method uses historical information to decide how much weight to give more recent history, trend or seasonality. These are the three main factors that you want to spend some time reviewing as they will give you some insight into the current demand pattern. The math will tell you how important each of these is, but it is up to people that understand the market to explain why.

This type of modeling is very helpful as it is data driven. I remember a meeting where a person on the marketing team mentioned that a certain product family was seasonal and after reviewing the data, it turned out that this was not statistically true. After digging into the details, it turned out that some products in that category had a strong seasonality component, but at the aggregate level you could not distinguish that pattern.

Another way to predict the future is to tag demand to macroeconomic variables and see if there is some correlation between them. A good indicator used in the plumbing industry is housing starts since it is a leading indicator of economic activity that tells you how many more homes are being built. There are a lot of different economic indicators out there and probably one of them is a good fit for your industry. Finding this correlation would be very useful since there are public forecasts available that can be used to predict the demand of your products.

Underlying Assumptions & Caveats

“History never repeats itself, but it does often rhyme” – Mark Twain.

What happens in an environment where there is volatility, uncertainty and complexity? Well, some time is spent discussing the data but this is usually not enough to understand all the caveats and assumptions.

One of the most important principles in forecasting is the underlying assumption that history will repeat itself and that past information can provide a good representation of what will happen next. The math tries to find certain patterns hiding in the data, like determining if the recent past is more useful or if there is a trend or seasonality involved. Early on in my career, when forecasting a high-volume item, we got a very high forecast for the near future, which did not made sense. It turned out that we had a big promotion in the recent past that was skewing the numbers. Once we took this out, the forecast corrected itself. The lesson we learned was that it is very important to scrub history to find outliers in the data.

Another important assumption is that external factors in the environment will remain constant, allowing the forecast to be developed under similar circumstances every time. Variations in the industry outlook, regulatory changes or economic growth fluctuations provide a challenge to the models described above. An alternate way to incorporate this information into the demand plan is required in these cases. A recent example of this is that after years of economic growth, the economy is stagnating or even shrinking. While this is known and discussed in the news and at the watercooler, there is always a lag from when this starts to happen and when the forecasting model picks it up.

Finally, a few important factors to consider in the assumptions are the horizon that you are using to plan and the level of aggregation. Think about the weather for a moment – usually the forecast for tomorrow is very accurate but looking at any day next month is not worth it. The same is true for any type of industry forecast; the further your look out, the less accurate it becomes.

It is a similar situation for the level of aggregation. It might be easy to predict how much you will sell in dollars for next month but it is harder to determine how many dollars per SKU or per customer you will sell next month. This is important to understand, because I have faced situations where the business asks how much we will sell in five years at SKU level by customer. It does not make sense to do a detailed analysis in this situation.  It is better to provide a directional number that is easy to explain and that supports the business to make the right decisions.

Measuring The Output Of The Process

Some of the most popular phrases in demand planning are “If I could predict the future, I would go to Vegas” or “The forecast is always wrong” and, my favorite one, “My crystal ball is broken”.  It is a fact that no forecast will be exact, but you can get within a decent range and with a good level of confidence. This is why it is good practice to measure the accuracy of your statistical forecast to understand the reasons behind your top misses.

From my perspective, it is important to understand how good the forecast accuracy is in terms of mix and volume. A good way of measuring mix is through MAPE (mean absolute percentage error), which basically tells you by how much you missed the forecast – regardless if you missed up or down – compared to what your actuals were. The advantage of this metric is that it is easy to understand (since it is a percentage) and it does not net out negative and positive values as it uses an absolute error.

The way to measure volume is through understanding if there is a bias at the aggregated level of your forecast. For example, when aggregating all your forecasts at the total dollar level by month, if you find out that the actual sales number has been below the forecast for some time, you might want to understand the reasons behind this. Ideally you want to oscillate between being a little above and then the following period a little below over the time horizon.

Discussing The Forecast During The Demand Planning Meeting

After understanding how a computer (or a Demand Planner) creates a forecast, it should not be a surprise why this needs to be reviewed in a group setting. Even if math is not your strength, a lot of value will come from having a thorough discussion of the historical data, the process used to come up with the numbers, the actual forecast, and metrics.

Below is a checklist of things to review. During this discussion, make sure that you balance time between the items that will move the needle, but covering in enough detail that allows you to understand the current situation.

  • Start with the input. Are we using the right dataset and have we looked at history to review and scrub?
  • How much importance do we give to more recent history versus the past? Is there a pattern in the data to be discerned, like trend or seasonality?
  • Is there an economic indicator that we could peg to a group or family of SKUs that will help us determine what the future holds?
  • Is there any external factor that could deter us from using our forecasting methods to determine the future values? Examples of this are changes in price (by us or the competition), changes in the economic or regulatory environment, or even new products that could cannibalize current demand.
  • What are the forecast accuracy metrics telling us? Are we better than last month? What are the top offenders and why?

In conclusion, a good statistical forecast provides a good start for your demand planning process and a solid foundation. The data, assumptions and metrics need to be understood and discussed in depth. But keep in mind that this only the beginning of the process. There is an art component of the demand planning process that incorporates changes in the environment, market intelligence and in general a consensus between areas that should be aligned with the company strategy.