Subscribe to DSC Newsletter

What methods would be most appropriate in trying to forecast when a directional change is likely to occur (in customers or revenue)?  Which are most sensitive to the early signs of a change?

My company is making a strategic change that will result in one group of customers (and associated revenue) going down while another goes up.  I have a regression-heavy predictive model that is hitting revenue and customers within +/-2% 3 months out, but this strategic change is bound do break that.  How do I spot the change in advance so that I can incorporate it into my predictions?  

Tags: forecasting, prediction, techniques, trends

Views: 149

Reply to This

Replies to This Discussion

I assume that your prediction model is modeled upon your historical data, therefore unlikely to be able to predict when and what the impact of a future unknown event. The best you can do is to build several "scenario forecasts" off your base forecast (i.e., best-case scenario, worse-case scenario, etc.) based on your judgment - The scenarios can be quick-and-dirty Excel calculations. The goal is to document your thinking process to the management. BTW, this is a common risk-management practice.

If you have an anlog dataset, that is, data from another company which has had similar event happening to them, then you can build a model to learn from their experience, and use the resulting coefficient to adjust your own predictions.

Unfortunately no analog; Just actuals going back for up to 4.5 years. My best idea so far is a time series model to smooth the actuals and then try to detect directional changes that are different from the historic seasonal trends. Still, that is unlikely to catch something more than a month out (ie I'll need to see at least a little change indirection before I can predict anything substantial).

Times series modeling would be useless for long-term forecasting (rule of thumb: 5 times of historical data to predict confidently 1 time ahead-forecasts). If you only have your own data to rely on to build your model, and you are interested in only detecting the "trend" in the short run, then try time-series modeling.

For abnormal, catastrophic event, no data-based modeling can be relied upon to predict the outcome. On a second thought, maybe a little Baysian decision analysis would help, where you inject a bit of expert judgment of the probability of the event happening. I am no expert on Baysiandecision science.
You are lucky in that at least this is a change that you know about. If you think you know which customer groups revenue will go down and which will go up, why don't you try simulating your model on newly constructed historical artificial data (e.g. revenue +/- 10%) which at least will give you an estimate of what to expect? If your simulated results are way off, that would mean your models sensitivity is off.

-Ralph Winters

If I understand you correctly, you should consider using a time trend variable.
Our software, Autobox, will automatically look for time trends to adapt to the data. This is done using "transfer function modeling" which as you may know is also confused with regression. Regression is estimate/forecast where TF modeling is IDENTIFY, estimate/forecast along with ARIMA and dummies.

Autobox also is on the look out for level shifts, pulses and seasonal pulses.

We would be happy to run through with a test dataset and send you back some results. You can even specify how you want the problem "bucketed(ie trend ---no trend)to be handled as we can override the Autobox defaults to customize the solution for your needs.

visit for more or call me at 215-675-0652

Tom Reilly


On Data Science Central

© 2021   TechTarget, Inc.   Powered by

Badges  |  Report an Issue  |  Privacy Policy  |  Terms of Service