Question: Which Statistical Forecasting Methods Should I Use?
Answer: Use the method that works with your data in your environment.
That’s it. Done. Next question. But that is too glib of an answer.
To elaborate a little bit, not all statistical forecasting methods work with all data. If your data is particularly unforecastable, perhaps your best bet is to use a simple 3-period average method. If you are using a system to do statistical forecasting, then you can only use the methods that are available to you inside that system. Some systems open certain methods for use only if you pay extra license price. So, you might be restricted by that. Some planners only prefer to use those methods that they are familiar with. If you are one of those, then your options may be limited.
The commonly used forecast methods can be grouped into 5 groups. In my previous blog, I provided some explanation of these groups and some examples.
Here are some general guidelines on selecting the right statistical forecasting methods for your business.
- Fit measures should not be used as sole criteria to pick the best method: Fit measures like R-squared and correlation only tell us how good the forecast fit the past. If fitting a forecast were the goal, why bother forecasting at all? Just say forecast = actuals in past months. This will give a perfect fit. The fit measures should not be used as sole criteria to pick the best method. Instead use some measure of forecast accuracy using hold out period. This implies that if you have say, 40 periods of history, do the forecast using the first 36 periods and see how bad a job they do in the last 4 periods. Once you select the best method, make sure to come back around and forecast using the entire history range.
- Segment Data: Different segments of the data might be better suited to different forecast methods. As expressed in the blog, it is advisable to understand what methods to use in which segment.
- Combine Methods: Consider methods that combine one or more effects (trend and seasonality for example).
- Only consider those methods that improve the forecast over the very simple methods. At the same time, be realistic. Do not expect magic. A 5% improvement over the simple methods might already be too difficult a goal to achieve purely by statistics.
- Pay special attention to fringe segments: Too little demand, too much variability, end of lifecycle. Ensuring these get the right method assigned will greatly improve the overall metric.
- Know your forecasting purpose. Use this purpose to derive the right level at which you need the forecast. Once this is done, select the best fit based on how good a job a method is doing at this particular level.
- Experiment with a lot of different models and pick what works. Try and avoid a ‘nervous system’ however, which changes the forecast every day.
- Try to understand a probabilistic range where the forecast is valid. Any point forecast generated by any method is guaranteed to be wrong. Methods that publish the forecast as a range around the expected value might do a better job overall.
So there you have it.