Humans are unpredictable
Forecasters do their best. They use complex mathematical models, they fit the data with the best technique. They consider the largest possible number of variables. But remember: no one can predict the human behavior.
We find tendency, the most likely way. But no prediction is perfect. Do not trust blindly any data.
Knowing the future you already changed it
Once you know what may happen, there is a tendency to: go to that way or avoid it. This is an instinctive natural process. Actuality, we work all day long expecting to be wrong. Our job is help our clients to positively change our forecasts.
When selecting your data to be analyzed, avoid exceptions and periods of not normal behavior.
Unless you really want to:
- Analyze your data considering drastic variations; or
- Predict the same behavior considering a known seasonal peak (like Black Friday, Christmas or a massive campaign).
Keep in mind that discrepant data directly affect the results. Forecasts tend to be conservative regarding significant variations. So, if you select historical data with peaks, they will affect the whole average.
We do Time Series Analysis and you know the seasonality of your business. Consider using full periods based on the calendar, like weeks, months and quarters.
If you are basically doing the same campaigns and expects an already known volume, you can select exactly the same period in the past to be predicted. For example, analyze last January to predict the next one. Of course this does not work if you have made huge changes in your business and expect great differences. Use this technique only if you are in a relatively stable scenario.
Analyze a couple of months ago do predict a couple of months ahead makes total sense. Actually this is most common technique. You will use recent data and consider your growth curve.
Quantity is not quality
Myth: “the more historical data, the better the prediction“. Unnecessary data hinder more than help. Long periods may contain unwanted variations or inaccurate data.
Select specific periods regarding your analysis strategy.