dashtabs / forecasting-models

Analysis and forecasting of macroeconomic indicators of the US gross domestic product, the consumer price index and the Ukrainian producer price index using autoregressive methods, fuzzy rules and a neural network with a long short-term memory.

Forecasting models

The project covers the analysis and forecasting of individual macroeconomic indicators based on several preliminary data using statistical models (ARMA, ARIMA, AR+trend), fuzzy rules, and LSTM. Real data were used upon request to the financial information provider Bloomberg for analysis and forecasting. All data are monthly since August 2016.

Input data requirements

All data files are .csv. All data is a monthly time series. In the .csv files data is presented in rows with the most current data to the left and old data -- to the right. The first row in the tables is always the month in format Mmm-yy. This row is converted to a datetime index in the program.

• gdp.csv – one-dimensional dataset.
• gdp_multi.csv -- two-dimensional dataset to forecast real value againstvthe nominal one.
• CPI-USA.csv -- many-dimensional dataset to forecast CPI index values using many features after performing correlation analysis and removing irrelevant indexes.
• PPO.csv -- two-dimensional dataset.

Output examples

Data pre-processing

Time Series Decomposition	Autocorrelation plot	Correlation Matrix

Forecasts

LSTM model	AR(2) + trend model	Fuzzy rule-based forecast

AR(2)+trend model metrics

AIC	DW	R2	MAPE	Theil
-149.996	1.865	0.9034	0.1866	0.0874

Theoretical background coming soon.

Project Structure

Utils

• analyzer.py – contains a class for GDP data pre-processing; performs ADF-test, prints metrics.

• common.py – calculates Theil coefficient.

• visualizer.py – visualizes autocorrelation and partial autocorrelation plots, also time series and time series decomposition.

Other files

• cpi_us_trend.py – builds multi-step and single-step forecasting models for CPI data.

• gdp_us_auto.py – implements an approach using auto_arima.

• gdp_us_empiric – implements an approach with manual setting of model parameters.

• gdp_us_fuzzy.py – implements fuzzy time series models.

• gdp_us_lstm.py – implements the construction of the model using the LSTM.

• gdp_us_other.py – implements a model with a step-by-step forecast with model retraining.

• ppo_ua.py – implements the construction of a model and forecast for PPO-data.