Supply Chain optimization for manufacturers (Deepvu)

Is the market steeling your profits? Look no further, we have a solution for you!

The best technologies are the ones that blend into everyday life. Since the invention of steel, we have been able to develop more technologies ranging from various tools (self-defense, basic utility) to building beautiful, massive infrastructures. As a consumer, steel becomes part of the background, but what happens when we stop paying attention to its value?

The market for manufacturing goods has always worked backstage, providing the commodities necessary to build the technologies we use daily. When the price of these goods fluctuate, so do our devices, changing the direction of our supply and demand curve and ultimately how our economy shifts. It’s important that we control how these prices change by knowing what their value will be several weeks in advance. In our project, we worked with the company DeepVu, focusing on steel, one of the most historic and integral inventions and commodities ever created.

DeepVu is a deep-learning startup focused on optimizing supply chains for manufacturers. They work with tier-1 manufacturers in both the US and Asia. One of the secondary sub-use-cases that they encounter involves forecasting the price of certain commodities that are key constituents of their customer’s products Bill of Materials.

For our project, we looked at 10 years of steel pricing data to build a Machine Learning model in order to predict steel prices up to 10 weeks in the future. Predicting future prices is a type of time series problems, so our model had to take into account both chronological and parametrical aspects, dealing with uncertainty due to the features.

We manipulated the dataset given to us and created a new one that contained for each week t, all features from the week, 2765 features, and all price averages for the next 10 weeks (t+1, …, t+10). Doing PCA, we then selected 50 features from this new dataset. After, we trained it using linear regression and random forest, and ended up with a root mean squared error of 30.

We then tried a deep approach using a Recurrent Neural Network, both with and without doing PCA beforehand, and noticed slightly better results when the PCA was used(root mean square error of 25), showing better results than the previous approach. We concluded that the RNN is better because it has a lower RMSE, although it remains noisy.

Though our model is not perfect, it’s a great starting point for such a complicated and challenging task. We hope that with more data and better detailed features, we can further analyze and predict the prices for this extremely important commodity.