We wondered if LSTM performances could be matched by a model focused on approximating the data dynamics. In addition, such approach could be used in other domains [e.g: predicting the behaviors of a set of molecular reactions (molecular engineering), as well as, providing a new approach to climate modeling]
What it does
Based on the data gathered, we built a model using historic data and macroeconomic data such as GDP growth rates, inflation rates, and the price of upstream materials to predict the steel price over on long time range. Given the provided time, we implemented 2 predictive algorithms for steel prices.
How I built it
We have scraped most of the data relevant to changes in steel price and processed our data to build a model. We have gathered 9 main analytics and built 2 neural network predictive model using LSTM and ODEnet.
Challenge I ran into
We started by implementing ODEnet on tensor flow but it turned out too challenging under such time constraints. Had to learn PyTorch in order to use the author source code. Standard ODEnet alone lacked past memories information and showed poor results, pushing us to switch to a combination with RNN and linear model, a combination suggested by the paper author themselves.
Accomplishments that I’m proud of
We achieved short-range prediction with good accuracy but our model falls short of prediction capability quite fast. We are proud of the data that can be found in our repository which we found, of building the neural networks with strong constraints. We are very proud of our innovative idea on a subject that is tackled over the world.
What I learned
We learned teamwork is crucial under a time constraint and learned to fast development under pressure.
While this work initiates some novel approach in tackling financial noisy time series data, it does lack a second analysis that we would have loved to put in place. Indeed it would have been interesting to look at the effect of sudden important variation in one of the feature. Our prediction is that the Odenet, approximating in a sense the real network dynamic, could highlight interesting analytical effects to this regard.