Reinforce seasonal forecasting potentials with AI
The first solution decreased the safety stock by 2000 tons of 10 seasonal products leading towards a cost reduction of 950.000 € per year.
Budget planning and demand forecasting is the most important steering instrument within supply chain planning. Nevertheless, it is an evergreen challenge to predict future demands accurately. The complexity driver is the huge size of portfolio forecasts which need to be generated each month. Most companies struggle with 1 to 15 million forecasts for each product, country and customer combinations. An accurate and robust forecast algorithm is mandatory to tackle this challenge. Especially, seasonality is a major driver for volatility and therefore a primary driver for supply chain costs.
It turned out, that the usage of Neural Networks dominates traditional heuristics in the detection of seasonal patterns. In contrast, conventional forecasting algorithms outperform Neural Networks in forecast accuracy, if no more than order history is used for modeling. Nowadays, combining Neural Networks with conventional forecast algorithms is key for tackling challenges in demand forecasting successfully.
The right process
We tackle the challenge of steering demand forecasting by combining AI and traditional methods in the best way in terms of statistical evidence. It turns out, that the following forecasting procedure is recommended:
- Apply Neural Networks to classify historical order time series to identify seasonal products.
- Assign suitable forecast models to seasonal products (e.g. seasonal regression, exponential smoothing).
- Create an ex-post forecast to approximate forecast accuracy for seasonal products.
- Review and adjust models and parameterization for high error and high demand volume products.
- Segment products based on importance and forecast accuracy.
- Derive an optimal supply chain and system configuration for each segment.
What we published so far: self-learning demand pattern detection
We trained an artificial neural network (ANN) on a dataset, which roughly contains 60 thousand already correctly classified time series with 36 demand periods (see Figure 1 left). The task for the neural network is to classify the time series correctly into the six different patterns. We trained the neural network on 32 thousand time series. 8 thousand were used for validation during the training. After the training phase, we used the neural network to classify 20 thousand times series, which were at this point completely unknown to the network.
Fig.1: Test and evaluate demand pattern detection
On the test-dataset the CNN reached an accuracy of 99.94%. 11 demand patterns were wrong classified out of 20 thousand time series (see Figure 1 right). The calculation for the classification takes 0.43 seconds, which is 0.0000215 seconds per time series. Compared to human or statistical the neural network wins – both in computation time and in accuracy.
Example: seasonal demand Forecasting
We use a product portfolio which consists of 5000 products. We provide for each product a demand history 40 month and we use one year for testing our ex-post forecast. Moreover, we compare three forecast models with the expanding window method: A first order exponential smoothing model (additive with alpha = 0.3) is used as reference model, a third order exponential smoothing model (additive with alpha = 0.3 and gamma = 0.3) without a trend component and a ANN based forecast model, which was trained on 5000 observations using the training horizon. Each model follows the principle of rolling forecasting. They are trained on the pass-dependent train horizon and tested on the next 12 months. The lag in each evaluation scenario equals one month (see Figure 2).
Fig.2: The forecasted values of all models
Table 1 summarizes the cross-evaluation of the models. In all passes, the third order exp. smoothing model dominates in terms of the measures MSE and WMAPE, followed by the Neural Network model. Both models induce a lower MSE and WMAPE than the first order exponential smoothing reference model. Thus, the third order exponential smoothing beats the ANN in seasonal forecasting in terms of accuracy.
Table 1: Summary cross-validation with expanding window method
Value case: decrease safety stock cost for seasonal products
We applied the suggested forecasting concept in a project within the chemical industry. The portfolio consists of 5200 products containing three full years demand history. First, we applied the Neural Network for demand pattern classification. It detected ten seasonal products with a total order volume of 17630 tons in the last year. In past, this task was executed manually: Seasonal products where missed or even wrong detected. Second, we used the third order exponential smoothing model to predict each product and compared it against the customer’s forecast. The outcome of this new forecast model allocation induces an increase of the forecast accuracy by 20.06%. This effect leads towards a safety stock reduction of 2000 tons and finally to a cost reduction of 950.000 € per year.
Take the robot out of the human
One aspect of digitalization is the reduction of error-prone manual tasks: Spending days of monotone work in solving monthly, weekly or even daily returning tasks will decrease human efficiency and “bore out” employees. It turns out that these tasks occur in many business areas such as demand planning. A clever mix of AI and traditional methods support a planner in finding demand patterns, predicting future demands and helps to focus on and prioritized important planning decisions. Thus, parametrization is almost fully automated, and the manual planning effort is reduced to a minimum. We build bridges between innovative and operative processes allowing to include the deep data insights in daily decisions. That is what Camelot’s digitalization strategy is about: Take the robot out of the human.
We would like to thank Torben Hügens for his valuable contribution to this article.