1
$\begingroup$

Let's say we want to forecast revenue by month for the next 12 months, and we have daily revenue data for the last 3 years.

We could then group this data by month, train our model using revenue by month, and forecast for the next 12 months. Alternatively, we could train our model using daily revenue data, forecast for the next 365 days and group the predictions by month to obtain a revenue by month forecast.

Which one generates more accurate results, performance notwithstanding? I wonder if there is a definitive, general answer to this.

$\endgroup$
  • $\begingroup$ Forecasts tend to become less and less reliable as you forecast further into the future. On that basis alone, I would think that forecasting over the nest 12 months would be preferred over forecasting over the next 365 days. But I may be off base with my intuition here. $\endgroup$ – Isabella Ghement Dec 13 '18 at 17:12
  • $\begingroup$ 12 months out and 365 is...the same. $\endgroup$ – Tom Reilly Dec 13 '18 at 19:54
  • 2
    $\begingroup$ @Isabella I find it intuitively helpful to contemplate data patterns that might be masked by the monthly grouping. Suppose, e.g., that my company recognizes revenue on its books only on Mondays. Such a clear daily pattern can be accurately forecast for a long time, but if the data are grouped by months then the analysis will have to cope with spikes occurring in the five-Monday months and seems unlikely to be able to forecast when such spikes occur in the future (based on just 36 months of data). Blindly modeling monthly "pulses" and "seasonal pulses" won't work well in such cases, either. $\endgroup$ – whuber Dec 13 '18 at 20:58
  • 2
    $\begingroup$ (Continued) In another direction, I have encountered financial time series where forecasting was improved through quarterly aggregation. The reason (it turns out) is that large investors were trying to meet targets at the end of each quarter, resulting in pushing some activity into the ends of the quarters that otherwise would naturally occur at the beginnings of the next ones. This made monthly forecasting difficult, with high prediction errors, but quarterly forecasting was reliable. To conclude, a general answer to this question likely eludes us, because it must depend on circumstances. $\endgroup$ – whuber Dec 13 '18 at 21:01
  • $\begingroup$ How hard is it to try both? $\endgroup$ – Reinstate Monica Dec 13 '18 at 22:31
1
$\begingroup$

It's always easier to forecast summarized data. Use the monthly data. You would want to use the daily data if you need a daily forecast.

$\endgroup$
  • 3
    $\begingroup$ One can only agree that using less data tends to be easier, but this question asks about accuracy, calling into question your recommendation. My experience has been that accuracy in forecasting can be lost by grouping data before analysis. Indeed, there's a theorem to that effect, but it should be qualified with the understanding that even if using all the data is potentially more accurate, one has to adopt a model and appropriate statistical procedure to actually make use of that information! $\endgroup$ – whuber Dec 13 '18 at 19:57
  • $\begingroup$ Modeling the data at a daily level would be more accurate...we agree....but who can model the data well at a daily level? It's quite complicated to do well so that means monthly wins....imho. lol about the seasonal pulse comment. It is not blind and is impactful. Do you want to have a sidebar to go over this together? $\endgroup$ – Tom Reilly Dec 14 '18 at 13:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.