5 ways how AI can improve your demand forecasting model

The concept of demand forecasting is nothing new in the ecommerce and retail industry. Often performed manually by analysts with the use of ERP software, it helps reduce out of stock levels, increase efficiency and improve customer satisfaction. However, with the advancements in artificial intelligence, demand forecasting can be brought to the next level that will give your company an additional edge. Using our project experience and AI-expertise, here are 5 key ways that deepsense.ai leverages AI to improve the accuracy of your organization’s demand forecasting model.

1. We feed AI algorithms with data sources including sales data, product characteristics and marketing campaign plans.

Traditional demand forecasting systems typically rely on the extrapolation of historical sales information, taking into account seasonal variations or the overlay of strategic initiatives (such as that the company will push product X next month). Unlike humans, AI can manually aggregate numerous data sources and produce meaningful results. In a project with an FMCG that carried out a lot of promotional activities, we leveraged external databases and marketing data. All of those pieces of information were combined together with historical data to produce a very precise forecast.

Combining data from various sources

The power of AI lies in the ability to process large amounts of information, which enables us to use multiple sources of data to feed our models. In the case of demand forecasting, the most obvious data to use is historical sales, i.e. the amount of items sold daily and their price. But forecasting demand is a multi-dimensional problem and many important factors influence demand.

Where to look for more data?

The darkest place is under the candlestick, isn’t it? Look at your own internal data. It contains  sales data, product characteristics, metadata, data on promotions and your marketing activities. External sources of data are also important, and include sales data from your distributors or even from points of sales or market reports provided by companies like Nielsen. – these are potentially excellent data sources. A third group to look at is contextual data, such as will emerge from a calendar of holidays, demographic or geographic data. It’s all there, you just have to turn over the right stones.

Internal data as low-hanging fruit

Probably the most valuable internal data in demand forecasting comes from promotions and marketing. You plan marketing activities and promotions, so the data is there, it’s just a matter of plugging suitable data sources into a demand forecasting system. It’s intuitive that low prices will increase demand, but by how much? You can figure that out by either analyzing historical data or, more effectively, feeding a demand forecasting system with promotions and marketing data to do the job for you.  Is lowering the price below a specific level an unwise step, if demand and revenue aren’t going to increase as a result? To find out, you can ask your demand forecasting model – just run a prediction for a hypothetical promotion level and see if your demand, or even your revenue, increases. That’s right, a demand forecasting system can also act as a price elasticity tool!

Plugging product characteristics and metadata into the system will enable it to draw on historical data of similar products while building a forecast. If you’re selling online, tracking the traffic on your website is also a valuable step, as doing so may allow you to uncover trends in your demand time series.

Which external data do you really need?

If you’re going to use external data, choose it wisely, as you’ll have to pay for it. Some data types are virtually always helpful, like sales data from distributors or points of sales. Sometimes increased demand from a distributor can influence demand from others, and it may not be easy to spot without an automated AI system. If you sell through distributors, such a system is one way for you to use the data you have at your fingertips.

Almost every business depends to some degree on the weather. While a daily weather forecast doesn’t influence total demand for warm shoes throughout the whole winter, it surely does influence daily demand for water in summer, especially in tourist areas. Weather data is almost always invaluable in demand forecasting systems. But it also requires caution:  because the weather changes daily, it can’t actually be used to train models. You want to use a weather forecast for the day you’re forecasting demand.

Demand is always context-specific

Demand is context-dependent for virtually every business. The same FMCG store can sell a very different amount of goods on the day before a “big” holiday and on a typical Tuesday. Although some conventional forecasting algorithms can handle changes in demand due to weekly seasonality, they can’t take into account the day before a “big” holiday. Features like “is today a holiday?” or “number of days to the next big holiday?” can easily be handled by AI models, which results in better forecasting.

What is a “big” holiday? That’s very society-specific. Demographic data helps define more contexts in which sales take place. Using it can obviously help you decide which holidays are “big” and which are not. But demographic data can give you far more than that. If you run a brick-and-mortar store, knowing the local population’s wealth can radically improve your forecasting in specific product categories – surely demand for top-shelf products will be much higher in wealthy areas. Geo data, meanwhile, allows you to use parameters like “number of similar stores in the neighborhood” or “distance to the nearest shopping mall.” They can tell you a lot about local demand, and also help you uncover patterns and dependencies that you wouldn’t expect to be there.

Do it the right way

The more data sources that are being used, the more intelligently they must be connected. Sometimes data requires a lot of preprocessing (because data is missing or the time resolution presents problems. For example, hourly weather forecasts will not be suitable for a daily demand forecast unless they are properly preprocessed). Sometimes the data is contaminated–there is a sudden change in units used or a lack of data, which manifests itself in strange patterns or constant values. Deciding which values you can actually trust and which you can’t is no trivial matter here. deepsense.ai’s experience has enabled us to come up with an effective strategy for using many data sources and selecting only the most important of them.

2. We use AI to predict demand for new items not yet present on the market

While not of interest for all industries, this feature can be crucial e.g., for the apparel sector. In a project for an e-commerce fashion retailer we used information about the garment characteristics including cut, color, pattern, texture, material, as well as photos of the garments. Finding similar items that were sold before, helped us assess how nuances (e.g., fabric) can impact sales. Our demand forecasting models then compared these characteristics to the attributes of previously sold items. We found that enabling customers to browse products that are not yet on sale is a simple yet powerful solution for predicting demand even more accurately, before products hit the shelves.

Demand for new items not yet present on the market

Demand forecasting in its simplest form is about predicting future sales based on historical sales of a given item. The more often you introduce new products to the market, the more crucial demand forecasting will be for new products. But for new products, there’s no historical data. Fortunately, the assumption that the product will behave similarly to comparable products that have been sold is reasonable. deepsense.ai developed a demand forecasting system for an apparel retailer for products the company had not yet made available for sale. It’s fast fashion scheme was at the heart of its business, and making better predictions about what might soon be a hit was crucial. But apart from by looking at them, how can we actually tell which products are similar to each other?

Extracting product characteristics

To find the similarities between products, metadata is essential. With apparel, such data include an item’s color, pattern and cut while in the food industry ingredients, brand, origin, flavour, package size and weight are key data. A picture of the item is also important, particularly in the apparel industry.

So which algorithms are used to work through all this data and photographs? It all depends on the model that is used. To predict a single value or to produce a single forecast for an initial sales period, the most powerful and popular algorithms are tree-based models. It’s of course possible to use other models, like MLP or even simple linear regression models.

The data scientists reading this article will know that it’s pointless to pass all pixel values of an image to a tree-based model. One solution is to feed it to a convolutional neural network like ResNet and extract a feature map from the last convolutional layer. This provides a low-dimensional information-rich representation of the image, which can be passed to the tree-based model. Metadata, which generally takes the form of categorical or numerical variables, can be easily passed to a tree-based model.

Product DNA

That’s all very helpful, but we can do even better. We don’t actually know whether vectors entering the model are similar for similar products, but it’s good when they are. Why? Because we can anticipate that demand for similar items should also be similar, so the input vectors to the forecasting model should also be close to each other. Fortunately, there are a few ways to obtain such vectors, or embeddings, as they’re known to data scientists – which will constitute specific product DNAs, and intelligently encode the most important product features.
One way to do that is by leveraging neural networks and sales information. The core of this approach is to train a neural network to predict which items are often bought together. From that network we can extract an encoded representation of the whole product. That is our product DNA. If two products are likely to be bought together, the dot product of their product DNAs will be close to 1, which means they’re similar.

Another method to obtain product DNAs is triplet loss for neural networks, designed directly to produce an encoded representation of the input. Triplet loss directly forces the networks to produce similar vectors for similar items, and vice versa, without the use of sales data. It’s based only on the product’s characteristics. The first approach is more valuable when the product DNA is supposed to reflect an intrinsic taste of the customer, while the second one is more product-oriented.
Embedding product characteristics in a single well-behaving vector (that is, one that has similar items with similar encodings) is a fascinating yet powerful method to enhance demand forecasting models.

Use the help of your customers

There is also a very powerful feature to use when products have not yet been put on sale. On their websites, companies can place a button customers click to request a notification when the product becomes available. The number of clicks and their distribution in time yield up essential information about demand. But you have to exercise caution with this feature, however: website layout and where the button is placed shouldn’t change much. Otherwise, this could change the meaning of the feature over time.

The impact of one of our solutions

One of the main goals of our project was to minimize out-of-stocks(a lack of product on hand), which obviously reduces sales figures. Our solution decreased out-of-stocks by around 30%. The graph on the left shows the situation before our model was introduced, and the one on the right, the situation after. The blue bars depict orders and green/red bars depict total sales amount.

3. With AI demand forecasting models we can easily predict the impact of promotions and marketing activities

An AI demand forecast model can do more than predict demand. It can also serve as a price elasticity tool. Due to its comprehensive nature, it can be used to test the effects of price changes on demand. By analyzing historical data, our AI-driven demand forecasting model can break down the demand into latent “real” component and the “promotional” effect. This knowledge is of paramount importance in the grocery or electronics industries, which run frequent promotions on specific items. With our AI-driven model, retailers can test how a given promotion will affect sales and hence pick the optimal price point.

4. We can also track anomalies in the observed demand levels

Our demand forecasting system can also help you monitor current activities, not just forecast for future ones. By integrating multiple data streams such as real time activity on your website, we offer advanced anomaly detection. In e-commerce, anomalies can be the result of technical glitches (which must be fixed quickly) or a natural spike in demand. As a result, you can start fixing the problem immediately or start analyzing the root cause of the anomaly. If you find a huge spike in face mask orders, in real time, perhaps you might want to place a substantial order with your distributor before it dawns on others to do the same, but for a higher price. Alternatively, with an advanced anomaly detection system, you should be able to avoid financial losses like the ones Foreo encountered when it sold ~40,000 devices for USD 9,99 instead of USD 279, before the technical glitch that was responsible was spotted.

5. We recommend AI for demand forecasting for its outsized impact and tangible results

Demand forecasting is all about precision: on one hand you don’t want to overstock and freeze up cash in excess inventory, but neither can you afford to miss sales opportunities. According to McKinsey, AI can reduce forecasting errors by over 20%. That’s a claim deepsense.ai can vouch for, as we delivered even more value in one of our projects. Working with a large CEE ecommerce retailer, we were able to reduce out-of-stocks by 30%. Depending on how your forecasting system is currently set-up, the gains could be even higher.

The benefits don’t stop there. From the very first day that your demand forecasting model is operational, your organization can track the improvement in key KPIs, including out-of-stock and inventory levels and the effectiveness of promotions. Additionally, the model can also measure “soft benefits” including fewer man hours spent on forecast analysis. These clear benefits stand in contrast to those offered by other recommendation engines, where attributing the real impact of the AI solution to the overall sales growth is challenging. With this in mind,  demand forecasting is a perfect example for the first AI use case for your organization. With clear and substantial impact you can convince stakeholders to invest in further AI-related projects.