With deep learning so much praised, you’ve probably read a lot about its positive side. Let's skip this "sweet part" and concentrate on the things that can go wrong with deep learning-based demand forecasts as well as offer some ways to overcome these problems.
1. Lacking sufficient data sets
Deep learning is not a magic 8-ball that produces results out of thin air. To create reliable forecasts, deep neural networks need your detailed sales data with such important markers as promotions, days of the week and holidays.
Ideally, DNNs require at least a two-year sales history to become adept at recognizing recurring demand patterns. However, if you can tolerate some occasional inaccuracy in predictions, you can settle for ‘good enough’ data sets, for instance, one-year sales history will do.
If you don’t keep track of additional sales details in your analytical system, a DNN can end up producing inaccurate demand forecasts. For instance, without having a promotion marker, you risk getting blown-up demand predictions, as a DNN won’t be able to distinguish between promo sales and occasional sales spikes.
How to solve: If your data looks quite different from the described ideal data profile, you can involve data scientists. They’ll build a test DNN, feed the available data to it and assess whether your data is of sufficient quantity to be converted into reliable demand forecasts.
And if you haven’t yet collected an adequate sales history, consider what data and with what degree of detail you’ll need and start collecting it.
2. Disregarding demand-influencing factors
A DNN won’t consider a demand-influencing factor unless you explicitly tell it to take this factor into account. You are probably curious what would happen if you missed a factor or two? Let’s consider two possible scenarios to see that.
- You run the chain of supermarkets and convenience stores and you disregard the store types factor while planning a DNN. Is it a big deal? No, it’s not. Although a DNN won’t differentiate the store types, this won’t dramatically hurt demand predictions for SKU X per store. For example, the DNN can still rely on yoghurt sales figures split by store, and it will not predict as huge a demand as 150 items of a certain yoghurt SKU for a convenience store because a c-store never saw such high sales of this yogurt.
- Now suppose that you sell clothes and the factor that you disregarded is the weather. Is it a big deal? Yes, it is! This blunder is very likely to hurt forecast accuracy. Ignorant of the weather’s influence, a DNN won’t see relevant dependencies. For example, unable to recognize that it’s getting colder a couple of weeks earlier, a DNN won’t adjust the demand for warm clothes.
How to solve: Your data scientists and subject-matter experts need to thoroughly analyze your data to capture all key demand-influencing factors.
3. Training that goes off track
Overfitting and underfitting are two kinds of problems that data scientists can face while training your DNN. Overfitting means that your DNN ‘learns’ the training data too well. As a result, it shows splendid performance when it relies on your historical data records but fails to produce any adequate forecasts based on the data it hasn’t ever seen before.
Underfitting means that your DNN fails to learn the relationships between the inputs and the relevant outputs in the training data. This means that a DNN is even incapable of building predictions based on the historical data, which is why it doesn’t go ‘live’ being unable to pass the training stage.
How to solve: In case of overfitting, you can expand the training data set, thus giving your DNN a chance to learn more patterns. On top of that, you can apply a dropout technique that deliberately ‘switches off’ some neurons to cut off some dependencies. You can also reconsider the DNN architecture to make it less complex, for example, by reducing the number of layers or the number of neurons in these layers.
In case of underfitting, reviewing the DNN architecture can help. However, this time you need to make it more complex. Underfitting can also be a sign of disregarding important demand-influencing factors, which is why your data scientists should double-check your factors.
4. Expecting that your employees will welcome DNNs with open arms
Believe me, they won’t. Most likely, they will be frightened by the introduction of DNNs rather than excited about it. What’s more, they can get a feeling that their experience and skills are of no value anymore because their managers choose to trust a strange bunch of algorithms over them.
How to solve: Organize a special meeting to introduce DNNs. Invite every department involved in demand forecasting: your Analytics Department, Finance, Sales, Marketing – everyone. During the meeting, clearly explain what a DNN is, how it works, and what benefits it promises. It’s also important to state that it won’t replace your employees and their expertise will still be extremely valuable.
Encourage a Q&A session to ensure that your employees grasped the concept. At the stage of DNN design, get your stakeholders involved. For example, you will desperately need your subject-matter experts at the stage of defining demand-influencing factors.
5. Believing that deep learning is omnipotent
Any demand forecasting tool has its limitations. As to DNNs, they are helpless when it comes to unforeseeable circumstances, such as natural disasters, strikes, slowdowns or government decisions. Besides, a DNN learns from your historical data and minds the demand-influencing factors you instructed it with. That’s why, if a rival store opens too close to yours or if a competitor starts an unexpected aggressive promotion (which will inevitably influence your demand), a DNN won’t be able to adjust demand forecasts accordingly.
How to solve: Bad news: nothing can really help when it comes to force majeure. Good news: when some external factors start to strongly influence DNN performance, you can do some fine-tuning and, say, introduce a new factor.
It’s nice to know the cure for each problem
Problems can arise at any stage – data preparation, analysis of factors, DNN training or corporate change management. Still, there’s a cure or at least a pain killer for each problem described. I hope that, with these tips, you’ll have realistic expectations about your deep learning-based demand forecasting efforts.