I write this post motivated with our recent development of new module in BusinessQ, intended for Predictive Analytics and Forecasting of data. In our Dashboard module for key performance indicators we are using heavily bullet charts. Very often, they show Actual data compared with budget and forecast values.
With building our predictive models we wanted to accomplish two things: first, to provide real forecasting for our bullet graphs and secondly, we want to expand our services and product capabilities with real predictions that with recognize patterns in historical data.
Small and Medium Businesses can benefit from Predictive Analytics in many ways. Predictive analytics is in a way business intelligence technology that produces a predictive score for each entity on organisation, being it a customer, a product , location etc. Assigning these predictive scores to entities is the job of a predictive model which has been trained over actual data. As a mathematical model we choose “Holt-Winters” model.
First, the actual data is prepared by cleansing it of abnormalities. We did this by introducing a factor for smoothing (or de-seasonalizing) the data. We are using a central moving median that uses a certain number of periods before and after a given period. A normal moving average will use only the periods before a given period to find the average. The effect is – the average will trail behind an event, no matter the number of periods used. A median on the other hand will cuts across the data where an average will show some sign of a spike. In that way, a median is a better choice to avoid abnormalities in the data. Note below how the moving average behaves with 3 different parameters.
Now we can put our predictive model in test. Basic test is to predict future based only on one-year data because we can easily check visually if model is working as expected. Actual data is agregated on monthly basis for easc entity (let’s say, a product).
This is the resul: (blue are actual data, orange is mathematical prediction):
Notice how pattern is well preserved, but it is taken into account upward or downward trend!
Great, now let’s see how the model will predict one year ahead based on three years of actual data:
Notice how Forecast model respects two facts: that product is every June and July (6,7) on every year sold much more then usually. But, we put last year actual very bad on purpose. The model rected in a way that it predicts relativly good August, but not so good as in first two years, because of bad last year sales!
Soon we will add this kind of predictive analytics in BusinessQ. For any more information, just contact us!