Forecasting prices for metal product

Winter School CompTech 2022
The project was carried out jointly with company www.evraz.com

Objective of the project

Based on the data accumulated over the past few years, build a predictive model to predict behavior market prices for metal product.

Out results

Prepared data:

• semantic analysis
• decomposition of data by meaning
• search for outliers
• filling in gaps
• generation of new features

Tried different models:

• linear regression
• random forest
• tree boosting
• fully connected neural networks
• recurrent neural networks

On the basis of a recurrent neural network, results of about 10% of MAPE were obtained. An application has been created that processes new user data and returns a prediction for the next week.

Repository structure

• preparcer: create many small tables of the usual format from the original large table, save them in tables/

• append_and_make_features: distributes new data across tables/, generates new features and dataframe-input for predictor models. It is important that tables/ exist and are non-empty!

• grid_lstm: training the lstm model on a grid of parameters in order to identify the optimal architecture

• lstm_predict: make a prediction of trained lstm

• docs/vision: detailed description of the project in Russian