We are witnessing plethora of information surrounding the impact of COVID-19 on supply chains and how it impacts the companies. As highlighted, there are very good use cases & papers of how different modeling algorithms work on normal circumstances like fast moving time series or usual casual methods. In case of intermittent demand we notice scenarios for products which are slow moving, spare parts of automobiles or products which at the end of the life cycle, however very limited scenarios with actual practical supply chain statistics are available for scenarios covering demand during epidemics or such events. Demand forecast is done based on the actual sales data or extrapolation of historical data, but for a situation like Covid -19 where the demand is sporadic and driven by social factors how would you forecast demand? Forecasting during epidemics is difficult problem due to the fact that the nature of the events are sporadic and location where the pandemic strikes. Once we are able to assess our forecasting ability for current situation, we would be able to test our systems for the worst forms uncertainty in our supply chain systems. What is needed in the current scenario is assessing the methods which can provide realistic predictions on demand which likely to arise from the frequent lockdowns or opening up the economies backed by human behavior.
Historically we all are aware of how Intermittent forecasting models such as Croston or ARIMA with Covariate can give you the forecasting during uneven demand patterns. With more and more companies are compelled to face this situation, what is needed now is complete evaluation of such models for ability to predict a specific segment of industry. The current trend of events and how companies are managing their inventory is crucial for the success of many organizations. Though Just In Time management of inventory has taken a backseat it does not implies that companies will have leisure of stockpiling the reserves. It needs a careful evaluation of current models and operational protocols. Still the old mantra of increasing the overall service levels and reducing the inventory levels for items makes sense. Companies needs to evaluate with wide range of forecast / inventory model combinations, and it is the opportunity for many to develop better models which can accurately predict difficult situations.
The current generation is fortunate to have accessibility to great Computing Powers and number crunching systems. Some of the enterprise systems are 4th generation powers offering excellent modelling capabilities Integrated Business planning systems using sophisticated machine learning algorithms. These algorithms offer accuracy during normal intermittent demand scenarios for all circumstances or assumptions like gamma factor. For example, in case of Croston Method, which is widely used forecasting methods for intermittent demand forecasting, two steps are used. Exponential smoothing and average demand calculation. The forecast is a combination of the estimated time lag between two consecutive positive demands and expected amounts of items sold.
When the lockout of similar measures artificially induced which do not follow a pattern of positive demands, how would you calculate the demand. Do you consider the local government norms of lockdown days or population or commodity type or items which needed for bare minimum survival? Such scenario is also challenged when you see the pattern of the products brought over by people during the lockdown. Some of the interesting articles published by some leading research firms like Mckinsey include Household supplies (Laundry, toiler paper, towels) followed by packaged food. Even there is a sudden stockpiling of medicines for common ailments. How would you model human psychology into demand forecasting? There are various approaches suggested for example of separating the events into one set of baseline and another one of extremes. The normal baseline can be analyzed using normal forecasting methods and special events which require intermittent forecasting. Separating the complete event into a set of baseline and special part is challenging to identify as highlighted earlier these events are sporadic and arise at different locations. What is needed in these conditions of epidemics is the health models which can gather the forthcoming epidemics and locations it can strike, which are not easy to build given the uncertainty in the exposure as we have seen with predictions carried out for USA case study.
The stockouts you have witnessed in May centers is not the real customer demand, and we need to rather understand the pattern of such demands which are induced by human psychology as well. When a forecasting is done with those sophisticated models, human judgement is also needed to assess the criterion. There is a good amount of literature which also supports the statement saying human judgement gives good forecasts. However, a combination of human judgement as well as statistical forecasts is also picking up demand in recent months. Like in case of Integrated Business planning, it can generate demand based on statistical forecast, which is then adjusted by human judgement to suit to the conditions. What is also needed in the current scenario is a model which can incorporate human behavior into forecasting models, again limited work has been done on this item. The current scenario is again a particularly good opportunity to understand the human behavior and demand driven because of the same.
To conclude it is the time for organizations to work with various models or mix of hybrid models and validate against their data which has been collected for past 5 months of time. Various models like Croston, ARIMA with Covariate, Exponential Smoothing with Covariate, Support Vector machines, Dynamic Linear Regression alone or in hybrid modes of models or along with human judgement may be tried to develop a scenario. Various performance factors may be assessed to identify the suitable forecasting method which can accurately predict the demand during this period. But definitely it is the time to assess the existing methods which are available in our arsenal for intermittent forecasting software system and carry out assessment with single model vs aggregate models vs behavior models or incorporating human judgement into models so as to identifying the model which best suits our requirement. Humans are not designed to live with problems, rather find innovative ways to circumvent the situations as well, and the current situation is an opportunity for companies to ideate strategies so that future generations are secure.