Even if you read data analytics news only occasionally, I’m pretty sure you’ve come across something called ‘deep learning,’ but how does it benefit demand forecasting, and why should you be using it for your demand planning?
1. Deep Learning Recognizes Both Linear & Non-Linear Dependencies
I won’t be surprised if your current demand forecasting solution assumes that we live in a perfect world where dependencies, if any, are exclusively linear. And I suppose you keep searching for different workarounds not to shatter this illusion. However, linear dependencies can’t describe your business environment. And here’s a simple example that perfectly proves it: If I told you that no matter the SKU or the category, a twofold increase in the granted discounts amount would always result in a twofold sales increase, you’d never believe me. And rightly so! A solid demand forecasting solution has to recognize non-linear dependencies to be able to consider the true impact of demand-influencing factors.
As for deep learning (and its dedicated tool – a deep neural network), it does extremely well at capturing non-linearity. In fact, DNNs’ architecture is specifically designed for it. It relies on activation functions that are inherently good at identifying both linear and non-linear dependencies.
2. Deep Learning Balances Storage Costs & Shortage Costs
The newsvendor problem is timeless. We must balance storage costs and shortage costs, with all the pain that it brings – either dissatisfied customers or extra warehouse expenses, electricity costs and disposals for perishable goods.
Can deep learning produce demand predictions that stop these balancing troubles? It absolutely can!
A DNN takes in your historical data. Ideally, these are your sales records for the past two years split by store and SKU with ‘promotion influence’, ‘perishable attribute’ and other markers. The output that the DNN will produce can be a demand forecast for SKU X (for the next week, store Y).
For a DNN to do its demand forecasting job, it should first be trained on your historical data. During such training, the DNN learns to find the dependencies between the inputs and the output. With that, you know what output it should produce and apply a dedicated technique – a loss function – to ‘tell the DNN off’ if it produces wrong predictions. This serves as an incentive for the DNN to adjust its weights and minimize its error. And besides that, loss function completes another important task – keeping storage and shortage costs in balance. In this case, the loss function uses the same principle – it ‘penalizes’ a DNN for the predictions with a wrong storage-shortage balance.
What’s more, you can create several DNN architectures and embed different loss functions there (with different storage-shortage ratios). For instance, you’ll definitely feel the advantages of such an approach while predicting demand for perishable and non-perishable goods.
3. Taking Advantage Of Data Diversity
Do you have a lot of ambitious ideas on how to forecast demand, taking into account diverse data like competitor activities and prices or trends derived from social media? If you don’t have deep learning to help you with that, you may have to settle for ‘piecing together’ some insights from different data sources and using some primitive tools and enormous efforts to put these insights together.
To turn such ideas into reality, DNNs process data from internal and external sources as well as taking in both numerical and categorical inputs, which expands your opportunities enormously. You can still feed your historical sales values to a DNN but additionally you can ‘tell’ it that those particular figures were promo sales on Friday in one of your supermarkets (all this additional information is categorical). This will let you find much more interesting dependencies in the data.
4. Deep Learning Keeps Your Decisions Unbiased & Based On Data
If your demand forecasts strongly depend on the opinion of procurement specialists who interpret data and make final decisions, you can’t be sure that your demand forecasts are totally unbiased. People tend to have some predefined concepts of what the demand can be, and when they look at data, they often unconsciously try to justify their ideas.
A DNN doesn’t know what it predicts – a weekly demand, an optimal inventory level or something else – so, it won’t play along with any assumptions. DNNs ingest data and expose it to weights as well as smart activation and loss functions that do the work. When the network is trained and data scientists verify that it produces accurate forecasts, DNNs stick to the parameters and logic that have proved to be right, no matter what new data it consumes.
Deep Learning Is worth Considering, Isn’t it?
If, after reading this blog post, it seems to you that deep learning may help you create reliable and accurate forecasts, I can consider my mission complete. But, by no means, am I saying that deep learning is the only option available. Undoubtedly, I find it appealing that DNNs can identify both linear and non-linear dependencies, weigh storage against shortage costs, analyze diverse data and avoid bias, but this technology is not a silver bullet and you should also consider the limitations of deep learning before deciding on what demand forecasting solution to opt for.