Forecasting demand is challenging. Various products exhibit different demand patterns. While demand may be constant and regular for one product, it may be sporadic for another, as well as when demand occurs, it may fluctuate significantly. Forecasting errors are costly and result in obsolete inventory or unsatisfied demand. Methods from statistics, machine learning, and deep learning have been used to predict such demand patterns. Nevertheless, it is not clear for what demand pattern, which algorithm would achieve the best forecast. Therefore, even today a large number of models are used to forecast on a test period. The model with the best result on the test period is used for the actual forecast. This approach is computationally and time intensive and, in most cases, uneconomical. In our paper we show the possibility to use a machine learning classification algorithm, which predicts the best possible model based on the characteristics of a time series. The approach was developed and evaluated on a dataset from a B2B-technical-retailer. The machine learning classification algorithm achieves a mean ROC-AUC of 89%, which emphasizes the skill of the model.
Forecasting demand is challenging. Various products exhibit different demand patterns. While demand may be constant and regular for one product, it may be sporadic for another, as well as when demand occurs, it may fluctuate significantly. Forecasting errors are costly and result in obsolete inventory or unsatisfied demand. Methods from statistics, machine learning, and deep learning have been used to predict such demand patterns. Nevertheless, it is not clear for what demand pattern, which algorithm would achieve the best forecast. Therefore, even today a large number of models are used to forecast on a test period. The model with the best result on the test period is used for the actual forecast. This approach is computationally and time intensive and, in most cases, uneconomical. In our paper we show the possibility to use a machine learning classification algorithm, which predicts the best possible model based on the characteristics of a time series. The approach was developed and evaluated on a dataset from a B2B-technical-retailer. The machine learning classification algorithm achieves a mean ROC-AUC of 89%, which emphasizes the skill of the model.
Forecasting intermittent and lumpy demand is challenging. Demand occurs only sporadically and, when it does, it can vary considerably. Forecast errors are costly, resulting in obsolescent stock or unmet demand. Methods from statistics, machine learning and deep learning have been used to predict such demand patterns. Traditional accuracy metrics are often employed to evaluate the forecasts, however these come with major drawbacks such as not taking horizontal and vertical shifts over the forecasting horizon into account, or indeed stock-keeping or opportunity costs. This results in a disadvantageous selection of methods in the context of intermittent and lumpy demand forecasts. In our study, we compare methods from statistics, machine learning and deep learning by applying a novel metric called Stock-keeping-oriented Prediction Error Costs (SPEC), which overcomes the drawbacks associated with traditional metrics. Taking the SPEC metric into account, the Croston algorithm achieves the best result, just ahead of a Long Short-Term Memory Neural Network.
Intermittent time series forecasting is a challenging task which still needs particular attention of researchers. The more unregularly events occur, the more difficult is it to predict them. With Croston’s approach in 1972 (1.Nr. 3:289–303), intermittence and demand of a time series were investigated the first time separately. He proposes an exponential smoothing in his attempt to generate a forecast which corresponds to the demand per period in average. Although this algorithm produces good results in the field of stock control, it does not capture the typical characteristics of intermittent time series within the final prediction. In this paper, we investigate a time series’ intermittence and demand individually, forecast the upcoming demand value and inter-demand interval length using recent machine learning algorithms, such as long-short-term-memories and light-gradient-boosting machines, and reassemble both information to generate a prediction which preserves the characteristics of an intermittent time series. We compare the results against Croston’s approach, as well as recent forecast procedures where no split is performed.
Demand forecasting intermittent time series is a challenging business problem. Companies have difficulties in forecasting this particular form of demand pattern. On the one hand, it is characterized by many non-demand periods and therefore classical statistical forecasting algorithms, such as ARIMA, only work to a limited extent. On the other hand, companies often cannot meet the requirements for good forecasting models, such as providing sufficient training data. The recent major advances of artificial intelligence in applications are largely based on transfer learning. In this paper, we investigate whether this method, originating from computer vision, can improve the forecasting quality of intermittent demand time series using deep learning models. Our empirical results show that, in total, transfer learning can reduce the mean square error by 65 percent. We also show that especially short (65 percent reduction) and medium long (91 percent reduction) time series benefit from this approach.