An independent statistical time series analysis has been conducted for each dealer separately in order to assess individual dealer performance. This analysis demonstrates wide variations between the forecast and actual order input results. These plots are based on monthly order input and there is extreme volatility. While the recent downturn is observed, it is clear from these plots that accurate forecasting by dealer cannot be performed with traditional time series analysis.
The plots of each dealer forecast can be accessed by the dropdown box. Randomly selecting dealers reveals a broad variation in final quarter predictions. The black line is the actual order input, the blue line is the forecasted order input, the red lines are the upper and lower bounds on the forecasts and the green dots are the underlying trend.
This interactive chart displays the projected order input for each dealer. The black line is the actual order input. The blue line is the forecast. The red lines indicate upper and lower bounds for the forecast and the green dots are the underlying trend. Select a dealer from the dropdown box to observe each dealers projections.
For the last quarter prediction the average error across all dealers, when taken individually, is 25% MAPE. Statistical time series in aggregate has a reasonable accuracy, but when predicting individual dealers performance the client needs to look to other techniques to achieve better accuracy.
Skilled time series practitioners can use extensive feature engineering to improve accuracy, but statistical time series with extensive feature engineering do not generalize well. The resulting poor generalization means each dealer will require unique feature engineering. Extensive feature engineeering for each dealer will require man-months of specialized and expensive individuals whose sole purpose is to create accurate forecasts.
The plots show that the later periods in the time sample have increasing errors. The red lines are the upper and lower error limits. The high and low predicted order input widen as the later periods are approached. Further, The final quarters in the forecast begin to demonstrate a single period lag. This demonstrates time series dependency on past data "catching up" in later periods.
A change point occurs when a underlying trend significantly changes direction. Statistical time series poorly predicts change points. As the change point becomes further in the past, the times series forecast becomes more reliable. The pronounced change point in the aggregate forecasts becomes less obvious in the individual forecasts.
Partner 1 demonstrates extreme volatility, but is forecasted for high performance in the next quarter. The high performancce is not achieved. A conclusion could be that if attention was directed to this potential high performer this failure may not have occurred, but the more accurate conclusion would be the forecast is faulty.
A time series is always useful to recognize historical trends and identify seasonality. This contributes to the overall understanding of the dealer's performance. The client's forecasting challenge is to improve these forecasts so that they may be useful in allocating resources and identifying targets for pruning or increased attention.