Statistical Forecasting :
Right or Wrong?
14 reasons why statistical methods may prove unsuitable:
-
Do the peaks/troughs happen at the same time each year? If not, how can a
statistical algorithm predict them? Characteristic of Sales and Marketing activity is
‘never the same thing twice’.
-
Statistical algorithms are typically slow to react to recent trend changes.
- Statistical algorithms are typically poor at forecasting ‘lumpy demand’ in
small customers.
- Statistical forecast is possibly good for long term trend. But how much of the
variation in month on month is attributable to this and how much to promos, competitor activity
etc.?
- Once you have your initial forecast, how much of the ongoing forecast
process is likely to be statistically based? Isn’t it the worst thing
potentially overwriting a forecast that has been carefully amended for promotions etc.
previously with a new statistical forecast?
- How much more accurate is one more month of actuals going to make the statistical
forecast? (Versus the improved accuracy that applying new management knowledge to
the existing forecast will bestow)?
-
In reality the main activity each month is not the generation of a new statistical
forecast. It is a review of the existing sales plan in the light of new information
– qualitative on events plus what was sold last month.
- How can a statistical forecast accurately react to forecast errors? Does an
oversale last month mean the start of an increasing rate of sale or merely business brought
forward from the next month? Does an undersale last month mean the start of a decline, or merely
that the stock pipeline was reduced and that actually there will be larger than expected orders
this month? It is surely impossible for the software to know. Yet actually the forecasts 1,
2, or at most 3 periods out are often considered the most business critical!
- Sales are the result of a complex cocktail of behavioural and synergistic factors which
cannot easily be quantified. For example, if your last promotion was very
successful in a customer, their buyer is likely to buy big and early on the next one. Whereas if
the previous promotion was unsuccessful they are likely to adopt the opposite buying strategy.
Two alternate perceptions of the same promotion! How do you reflect that in ‘clean’
and ‘dirty’ or however you wish to ‘clean up’ the data? (Is a stock-outalways a lost sale? Probably not – maybe the customer will order more next time,
or more of a different size/product this time. Again, judgement is required.)
- Statistical approaches like ‘cleaning past promos out of the history’ and
adding them back to a statistical forecast are naïve, overly-mechanistic and do not
remove the need for judgement. (They replace judgement and resource at the
forecasting end with judgement and resource at the cleaning up history end. Plus, how do
you accurately quantify the ‘standard’ rate of sale without significant
judgemental resource?)
- If you think statistical forecasting is suitable, prove it to yourself with a
subset of your data. Don’t just assume!
- Apply alternatives to statistical forecasting – forecast steady / low
sellers less frequently, or review by exception based on accuracy analysis. Focus time on big /
highly variable items.
- Apply the 80:20 rule to what you forecast.
- Forecasting aggregate groups of small customers is easier because there are fewer
numbers and the individual peaks and troughs tend to even each other out.
- Use 80:20 to identify the skus to apply the strongest forecasting focus to. Small or
steady selling lines should need much less forecasting time and their inventory is much
less costly to support.
- It is remarkable how many businesses do follow the 80:20 rule on both products and
customers. If you haven’t analysed your data in this way it is definitely worth a
go!
- Don’t forecast volumes in isolation of the wider business plan. Manage
volumes, revenues and margins within a single, integrated forecasting process. So that all parts
of the business buy into the same vision of the future and can work in a coordinated way to
deliver it.
Conclusion:
Unless you are happy to produce daily, weekly or monthly forecasts by product exclusively using
statistics and without manual adjustment, then the activity of generating more accurate forecasts is
potentially and unavoidably a time-intensive process. The more time you spend on it, the more likely
you are to generate the most accurate possible forecasts.
Please see our article on 'forecastability' - a way of identifying which combinations are most and least suited to statistical forecasting / machine learning.