Time Series Forecasting: Which Error Metric Should You Actually Use?

The short version

The uncomfortable truth about time series forecasting is that most organisations treat it as a tooling problem when it is, almost always, a decision-making problem.

The second-order effects matter more than the first. A change that improves one metric usually degrades another, and the teams that do this well are the ones who name the counter-metric before they start rather than after someone complains.

Where teams get it wrong

Start with the constraint rather than the ambition. In most Indian organisations the binding constraint is not talent or budget — it is the absence of a clear owner. Once a named person carries the outcome, the sequencing question tends to answer itself.

Finally, there is the question of who pays and who benefits. When those are different groups, adoption stalls regardless of how good the design is. Aligning them is unglamorous work and it is usually the work that matters.

A practical approach

The second-order effects matter more than the first. A change that improves one metric usually degrades another, and the teams that do this well are the ones who name the counter-metric before they start rather than after someone complains.

Start with the constraint rather than the ambition. In most Indian organisations the binding constraint is not talent or budget — it is the absence of a clear owner. Once a named person carries the outcome, the sequencing question tends to answer itself.

What to do next

Finally, there is the question of who pays and who benefits. When those are different groups, adoption stalls regardless of how good the design is. Aligning them is unglamorous work and it is usually the work that matters.