Advances in forecasting sporadic and intermittent demand products
14 May 2013
14 May 2013
Researchers at the Lancaster Centre for Forecasting have investigated ways to improve forecasting performance of intermittent or sporadic demand products.
Intermittent or sporadic demand appears when a product has several periods of zero demand. There is uncertainty in both the timing and the size of demand. This is common for spare parts and highly specialised durable goods, such as heavy machinery.
Intermittent demand is a very difficult forecasting problem. The special demand properties of such products make the use of conventional forecasting models problematic. In turn, this results in inaccurate forecasts and substantial over- or under-stocking and waste. This is particularly interesting when the products in question have short life-cycle and may become obsolete fast. Although there are specialised forecasting methods to deal with this problem, their application and use is neither simple nor always results in optimal predictions. If your organisation faces such problems, the Lancaster Centre for Forecasting offers training courses on intermittent demand to help you build your in-house expertise.
Researchers at the Centre for Marketing Analytics and Forecasting have explored alternative methods for overcoming problems associated with intermittent demand and improving forecasting accuracy, as well as automating the forecasting process.
One approach is based on the idea of aggregating the historic demand across time with the aim of transforming the intermittent time series into conventional ones. The resulting new series have limited or no periods of zero demand. This mitigates the forecasting problems associated with the intermittency of the demand. On the other hand, new forecasting modelling questions arise. For instance, an important question is related with identifying the optimal level of aggregation. Different alternatives were evaluated and best practices were developed, leading to improvements of 10% to 20% in terms of accuracy, depending on the base forecasting model that was used.
A separate stream of research explored the application of neural networks on intermittent demand problems. Such models have the advantage that they can capture potential nonlinearities in the data, in particular if there is some underlying connection between the observed demand sizes and intervals between demands. Empirical evidence suggests that this connection is common in practice; however, the models typically employed by companies are unable to capture this additional information, resulting in poor forecasts. The results of our investigation suggested that neural networks resulted in superior inventory performance with substantial savings.
With the development of increasingly complex forecasting models for intermittent demand time series new problems arise. The Lancaster Centre for Forecasting is currently exploring questions related to model selection and parameterisation for intermittent demand. How to pick the best model for your data? How to ensure that it provides the optimal forecasts? If you would like to know more about the research of LCF in the area of intermittent demand please contact Dr. Nikolaos Kourentzes or Dr. Fotios Petrpoulos. The centre is always keen to embark on new collaborations with organisations and industrial partners.
Nikolopoulos, K., Syntetos, A., Boylan, J., Petropoulos, F., & Assimakopoulos, V. (2011). An Aggregate - Disaggregate Intermittent Demand Approach (ADIDA) to Forecasting: An Empirical Proposition and Analysis. Journal of the Operational Research Society, 62(3), 544-554.
Spithourakis, G., Petropoulos, F., Nikolopoulos, K., & Assimakopoulos, V. (2013). A systemic view of the ADIDA framework. IMA Journal of Management Mathematics.
Kourentzes, N. (2013). Intermittent demand forecasts with neural networks. International Journal of Production Economics, 143(1), 198-206.