Forecasting for Business - Blog

Most engineers will tell you that:

You can't optimize what you don't measure

Turns out that forecasting is no exception. Measuring forecast accuracy is one of the few cornerstones of any forecasting technology.

A frequent misconception about accuracy measurement is that Lokad has to wait for the forecasts to become past, to finally compare the forecasts with what really happened.

Although, this approach works to some extend, it comes with severe drawbacks:

• It's painfully slow: a 6 months ahead forecast takes 6 months to be validated.
• It's very sensitive to overfitting.  Overfitting should not to be taken lightly, and it's one the few thing that is very likely to wreak havoc in your accuracy measurements.

Measuring the accuracy of delivered forecasts is a tough piece of work for us. Accuracy measurement accounts for roughly half of the complexity of our forecasting technology: the more advance the forecasting technology, the greater the need for robust accuracy measurements.

In particular, Lokad returns the forecast accuracy associated to every single forecast that we deliver (for example, our Excel-addin reports forecast accuracy). The metric used for accuracy measurement is the MAPE (Mean Absolute Percentage Error).

In order to compute an estimated accuracy, Lokad proceeds (roughly) through cross-validation tuned for time-series forecasts. Cross-validation is simpler than it sounds. If we consider a weekly forecast 10 weeks ahead with 3 years (aka 150 weeks) of history, then the cross-validation looks like:

1. Take the 1st week, forecast 10 weeks ahead, and compare results to original.
2. Take the 2 first weeks, forecast 10 weeks ahead, and compare.
3. Take the 3 first weeks, forecast 10 weeks ahead, and compare.
4. ...

The process is rather tedious, as we end-up recomputing forecasts about 150 times for only 3 years of history. Obviously, cross-validation screams for automation, and there is little hope to go through such a process without computer support. Yet, computers typically cost less than business forecast errors, and Lokad relies on cloud computing to deliver such high-intensive computations.

Attempts to "simplify" the process outlined are very likely to end-up with overfitting problems. We suggest to say very careful, as overfitting isn't a problem to be taken lightly. In doubts, stick to a complete cross-validation.

It has been a little while since we posted some technical content about our forecasting methods. Let's discuss further the case of promotion forecasts.

Promotion forecasts are especially interesting for two reasons:

• promotions represent a heavy part of the business for many companies.
• forecasts are likely to end up really wrong when ad-hoc methods are used.

In order to clarify the traps looming behind promotion forecasts, let's have a look at the following schema. The two black curves represent fictitious product sales - illustrating typical consumer response to a promotion.

Disclaimer: Those behaviors have been measured on Lokad databases. Yet, market responses vary a lot from one promotion to the next and from one business to the next. Our purpose here is not to provide accurate quantitative models, but rather to focus on what we believe to be key insights for promotion forecasts.

1. Primary promotion impact

The most important effect of a promotion is usually a large sales increase of the promoted product for the duration of the operation. This effect is illustrated with (1) in our schema.

Guessing that the sales are likely to increase is trivial, yet precisely estimating the impact of the promotion - that is to say the extra sales generated by the promotion itself - is a complicated process. Lokad is using its tags+events framework to do this.

It can be noted that classical forecasting methods such as exponential smoothing are completely missing promotional patterns.

2. Demand drop due to market saturation

The second effect of a promotion - which is too frequently overlooked - is the demand drop that comes just after the end of the promotion. See point (2) on the schema.

Indeed, by observing thousands of promotional operations, we have found that frequently, just after the end of a promotion, sales levels are dropping below the initial pre-promotion sales levels.

This drop reflects the temporary market saturation caused by the promotion itself. Basically, people who would have bought the product anyway have hurried their decision, resulting in fewer sales when the promotion stops.

When this drop is overlooked AND combined with a naive forecasting method, the combination can lead to extremely poor inventory management and overstocks. Indeed, the delayed behavior the exponential smoothing is likely to suggest inventory replenishments precisely when the demand is going to drop.

3. Mechanical echo due to customer synchronization

The third pattern comes from the synchronization effect of the promotion on customer replenishments. This point is illustrated with (3) on the schema.

All products (consumable or not) tend to have a lifecycle of their own. The promotion is synchronizing - to some extent - consumption patterns of customers. As a result, after the initial shock (the promotion itself), there is likely to be decreasing echos in customer demand.

In our measurement, it appears that most of the time only the first echo can be measured with enough confidence to be of any use for forecasting purposes. Subsequent echos do exist but they are too diffuse to be used to refine forecasts.

Again, if the demand bounce (3) happens right after the demand drop (2), then the exponential smoothing model is likely to cause frequent inventory shortage, because it completely fails to anticipate the bounce.

4. Demand drift due to market evolution

The fourth effect is the demand drift caused by the promotion. This effect is illustrated by point (4) in the schema.

A promotion acts as a market perturbation, and the base sales level is likely to be different after the promotion; hopefully higher, but our measures shows that the opposite, i.e. lower sales, also happens, yet with a lower frequency.

In our experience, establishing a quantitative forecast for the drift is really hard. Yet, the drift can be taken into account in order to speed-up forecasts adjustments after the end of the promotion.

5. Sales cannibalization

Sales cannibalization is typically the second major effect of a promotion in terms of absolute sales volumes. Cannibalized products end up with reduced sales during the promotion as illustrated by (5) on the schema.

Yet cannibalizations are harder to model, because the effect can be diffuse and potentially impact a wide range of unrelated products. Indeed, depending on the situation, some customers may have a fixed threshold for their spending. The immediate consequence of the threshold is that the opportunity spending offered by the promotion ends up compensated by restrictions on somewhat unrelated products.

Lokad is trying hard to provide fine grained promotion forecasts, taking into account all the effects that we can measure. If you decide to go for in-house forecasts (which we don't recommend, but well), we suggest to make sure that neither (2) nor (5) gets overlooked.

Forecasting is hard, even when a significant amount of historical data is available. When historical data is limited, forecasting is much harder. But then, what about forecasting when there is simply no historical data?

The no-data situation is more frequent that it looks: for every product launch, a company has to forecast future sales for the new product, while there are precisely no records for this product.

In practice, we have found that many companies - already using robust statistical tools to forecast their regular sales - just guesstimate when it comes to product launches (or one-shot promotions). We have also found that, in many situations, guesstimates are vastly inaccurate.

Obviously, if there is absolutely no historical data, then, indeed, statistical forecasting tools (such as Lokad) are powerless. Yet, in most companies, new products are launched on a regular basis, and this history of launches can be analyzed to figure out patterns of early sales.

Lokad takes advantage of historical product launches (when such data is available) to forecast the sales of a product even if there are no data yet for this particular product. Typically, we estimate that 20 product launches or so are needed to start learning launch patterns. In practice, there is no hard-coded lower limit on the number of product launches in our technology, but with less 20 launches, forecasts tend to become erratic.

In practice, you can use the Safety Stock Calculator to forecast product launches. Note that raw sales data is not enough in the case of product launches, tags and events are needed as well (well, at least tags or events):

• Tags should be provided in order to describe the product. Tags typically express similarities that exist between products (ex: color, size, category, product family, ...). Those tags are used by Lokad to match the new product with existing ones. Typically, a tag is a permanent descriptor of the product: it does not change over time.
• Events should be (eventually) provided to describe the launch operation itself. Events are just like tags, but positioned at a certain date. Events typically represent marketing operations that support the product launch. An event usually has a lifetime shorter than the product itself (otherwise it should be considered a tag).

The distinction between tags and events helps Lokad to figure the relative position of the product within the distribution channels of the company (tags), from the impact of the marketing operations themselves (events).

Still unsure how to proceed with your product launches? Don't hesitate to drop us questions in the forums.