Taxi_Demand_Time_Series

Taxi Demand Time Series

Description

The project is dedicated to predicting the taxi orders number time series.

The task is to regress the time series 1 step forward.

Target metric - RMSE.

Stack

Loading, storing and processing data: pandas, numpy, phik

Working with time series: statsmodels, pmdarima, prophet

Visualization: seaborn, matplotlib, built-in library methods for time series processing

Automation: sklearn, category_encoders

Generic models: sklearn linear regression, LightGBM, pytorch (architectures: dense, LSTM), lineartree

Individual solutions:

• automated feature engineering and selection for classical ML models
• a forecast was obtained for the entire test sample with models capable of predicting many steps ahead
• a hybrid boosting model was used, which combines the advantages of linear models (stability) and gradient boosting

Data

The dataset of Yandex Practicum:

• 26,000 time series records with dates
• no missing values and outliers

Conclusions

• trained all the main models used for forecasting time series, including:
  • exponential smoothing
  • linear regression on lagged features
  • gradient boosting on selected features
  • SARIMA
  • Prophet (ready-made library for working with time series)
  • fully connected neural network
  • LSTM
• it is shown that models specifically designed for time series, such as exponential smoothing and SARIMA, provide the best result
• at the same time, linear regression with a large number of lag variables also provides a great result when predicting 1 step ahead