The Hippocratic Oath is an oath of ethics historically taken by physicians. According to Wikipedia, it requires a physician to swear to uphold specific ethical standards. The original oath is rather long, but it is commonly simplified to this idea: “First do no harm.”
This got me thinking: Is there an equivalent for forecasters? What harm(s) can a forecaster easily avoid?
Based on my past work, here are some examples:
- Avoid the harm of being worse than the naïve (or simple) forecast.
What is a naïve forecast you ask? Here are some examples:
Forecast in this period
- = Actual in the previous period, OR
- = Average of actuals in previous 3 periods, OR
- = Actual in the period which happened exactly 1 cycle (most commonly a year) earlier
- All these are super easy and can be done in Excel. One might even consider taking a weighted average of these naïve methods and calling it the forecast.
- Stop creating unnecessary and avoidable errors by forecasting the unforecastable combinations. For example, only 1-2 data points in history, too big of gaps between data points, etc. Very often, users will extrapolate these minimal data points to some trend. The data simply does not exist to suggest a trend in many cases.
- Stop making the forecast worse by including the human input that is demonstrably bad. You will find that there are plenty of forecasters who consistently make the forecast 5-10-15% points worse compared to the naïve forecast. Ignore their input. This must be done properly though, after appropriate conversation backed by facts. One should avoid doing it in a way that creates an emotional response or a political backlash.
- Often this will include clear signs of bias. Bias, once identified, is easy to correct.
- Cultivate a way to combine the different forecast inputs that are proven to be good. This could be across product families, time periods, or people. For example:
- For product families 1, 3, and 7, the statistical forecast is the only one worth considering.
- For product families 2 and 5, sales input is much better.
- For product family, both forecasts are very good and therefore an average of the two makes sense.
- One can similarly differentiate across periods; in periods in the near future, sales input might be good. In the outer periods, their forecast might need to be completely ignored.
Agree? Do you have other ideas? Please share via comments.