Winter School CompTech 2022
The project was carried out jointly with company www.evraz.com
Based on the data accumulated over the past few years, build a predictive model to predict behavior market prices for metal product.
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.
-
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 acrosstables/, generates new features and dataframe-input for predictor models. It is important thattables/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