This is the first time we combine the ordinal and continuous features with the prior historical order input performance to find a reliable trend and changepoint forecast for each client dealer. The technique used here is a deep learning method referred to as recurrent neural networks, specifically, a long-short-term-memory ("LSTM") analysis.
The trend analysis demonstrates a remarkable 9.9% MAPE, computed by averaging across all dealers individual accuracy. Adding the features to the prediction method reduced the statistical time series error from 25% MAPE to about 10% MAPE for all the individual dealers. The following plot demonstrates the trend prediction for each dealer along with the statistical time series forecast.
This data table is interactive and displays the deep learning trend for each dealer with the red line. The deep learning forecast is blue and the time series forecast is green. By selecting a dealer from the drop down box a new dealer is displayed.
The underlying trend for each dealer is clearly revealed. Each dealer benefitted from the rising edge of the product life cycle, but began to decline at different times, with some lagging by as much as 4 quarters. The client's product potential has peaked in its original market and is now in decline.
Sample interpretations may improve the effectiveness of this visualization.
Partner 1 demonstrates high-performance with above-trend line performance forecasted in the next quarter. The sales efforts are still rewarding this partner with above-trend line performance. Comparing this dealer results with other dealers in the same class and region may reveal methods to this client that will sustain other similar dealers.
The forecast for this partner is almost exactly equal to the trend line. This is not an inspiring result since the trend line is a measure of average performance resulting from the feature data and and historical order input. This dealer may be trying to perform well, but the market is not responding. Sales and support focus by the client on the dealer may yield improved results.
Partner 62 is most likely operating as a distributor and floats with the market conditions. If a competitor would offer better market dynamics this dealer may bolt from the client.
Partner 473 is clearly in trouble. This partners activity, training and satisfaction have significantly declined along with the market and price indices. The trend line is definitely down. Near-term triage by this client is required. It is very likely that Partner 473 is the target of poaching by a competitor.
The deep learning method employed, i.e., LSTM, here significantly improved the forecasts. A 10% MAPE rate on the forecasts for each of the client's dealers renders the forecasts to be reliable for this client's decision-making regarding the distribution channel effectiveness, resource allocation and internal forecasting purposes.
The decline in the trend line revealed by this analysis is contrary to this client's desired results. However, the client's challenge to identify trending performance by dealer is met. The next step in the example provides the highest accuracy for short term forecasts by dealer.