- How SupplyPike uses machine learning to forecast sales
- SupplyPike’s deep learning models
- The difference between SupplyPike’s models and others
SupplyPike forecasting takes current retail forecasting to a new level. We use deep learning, a type of machine intelligence similar to and inspired by the human brain. Instead of relying upon classical, statistical, or linear methods that have weaknesses and inherent predispositions towards different aspects, our models have no biases.
These deep learning models are complex, composed of layers of “neurons” – small nonlinear equations called an artificial neural network. A single layer of these neurons is capable of modeling any data trend. The artificial neural network learns from the data presented to it rather than applying a specific equation.
Our models look at the entire history of time-series information available, typically two or more years of sales data when demand forecasting. They iteratively repeat examining time windows across the whole history of data until they can successfully understand and predict the historical data. We call this process backpropagation.
With backpropagation, we present the neural network with an example from the historical data set, ask it for its prediction/forecast, and then calculate how correct the prediction is. We tweak the network’s millions of neurons slightly, pushing the neural network just a little closer to the correct answer. We repeat this process millions, sometimes trillions of times, in random order. The neural network slowly begins to build a model of the data presented to it.
Layering in neural networks
The layering system is essential to deep learning. Raw data presented to the neural network flows through one layer at a time until the output or the prediction remains. Each layer learns to extract different details from the data, slowly turning it into information that the successive layers will collate, interpolate and expand upon before pushing the information to the next layer.
Each layer gives the following layer more information to make the final prediction. We call this process training. Once we’ve trained the neural network, we can further refine it as we find more data. Such a neural network examining time-series data can make highly intuitive and accurate forecasts.
Deep learning versus statistical models
Deep learning neural networks are much more flexible, accurate, and less biased than statistical methods like Seasonal ARIMA, Holt-Winters, and other models. Deep learning is a superset of these statistical methods. The neural network can learn to apply exponential smoothing, moving averages, and other techniques used by those models but not be limited by them.
Using a neural network for forecasting leads to less bias towards a specific implementation and reflects the historical data that trained the neural network. Also, a deep neural network is not limited by simple historical data input. The input and output can be complex and multi-variable.
At SupplyPike, we use this aspect of neural networks to make our demand forecasts as accurate as possible. We incorporate many different pieces of information from the supply chain to forecast the demand. We can include data pieces that others simply can’t because we have a comprehensive picture of the retail data for a product.
We give our neural networks extra data that we call “context,” that is, contextual information built from disparate retail and non-retail data sources. These data include store counts, pricing, weather, sentiment from online postings, custom events, holidays, and broader retail trends such as category-level trends, panic-buying, etc. The neural network takes this contextual information and learns how it correlates with historical sales.
The SupplyPike forecast is entirely automated; without any user input, we create dozens of forecasts weekly for each SKU our customers have. We require our models to be able to predict both the corporate-level forecast and store-level forecasts at the same time while using the contextual information we provided. Simultaneously predicting store-level and corporate demand means we get both the advantages of a top-down and bottom-up forecast approach.
We’re able to create both types of forecasts because the output from the neural network can be as complex as we desire. We store historical forecasts forever, analyzing them for further improvements to our current capabilities.
SupplyPike’s Retail Intelligence software takes machine learning and creates actionable insights and attractive, easy-to-digest metrics. The SupplyPike Sales Forecast clearly models and visualizes your data.
Schedule a tour today to see your data in action!