Flat forecasts

Published on by Joannes Vermorel.

Statistical forecasting is a counter-intuitive science. This was already said in the past, but we are going to emphasize again this point.

Frequently, we get people asking for support because they have just pushed some data to Lokad, and the forecasts they obtain are flat. In other words, the forecasted values are constant for all steps ahead. Ex: constant sales values for the next 6 months, if we are considering 6-month ahead monthly sales forecasts.

It's perfectly clear though that there is not-chance for business sales to be perfectly flat for the next 6 months, so why Lokad keeps producing such meaningless results?

Well, we know for sure that business is going to change (at least a little) during the next six months. No question about that. Yet, the problem is: how can we produce a forecast as close as possible to those future changes? If we take the statistical road, then we need a statistical model to the forecasts.

The problem is that we need a good forecasting model; and the cardinal rule of statistical forecasting is that the more complex the model, the more data is needed for the model to be reliable. Models producing distinct forecasts for each step ahead are definitively more complex than the ones producing the same value for all steps ahead.

The other way around, we can also say that those more complex models are also less reliable on limited datasets which means that using them is very likely to decrease the overall forecasting accuracy in certain situations.

Back to the situation where people complain about flat forecasts, what is usually happening is simply that the data that has just been uploaded is either very short (like only 3 months of monthly history) or very sparse (like an eCommerce with only a handful sales for each product). In those situations, Lokad frequently goes for flat forecasts.

It's not a bug, it's an accuracy-improvement feature.

Categories: accuracy, business, forecasting, insights, usability Tags: business flat forecasting history insights models sparse time-series