Prediction accuracy will be evaluated based on the root mean square logarithmic error:
![](https://storage.googleapis.com/kaggle-media/competitions/hhp/2496/media/eval-metric.png)
Where:
i is a product
n is the total number of products
p is the predicted sales volume (units sold in week 26)
a is the actual sales volume (units sold in week 26)
Note: the metric is calculated using the natural log.
We have chosen a log-based metric to minimise the effects of outliers and reward accurate predictions, without overly penalising predictions that are far from the actual.