regime-detection

Dataset:

Big Macroeconomic dataset from FRED St. Louis desiged by McCracken and Ng (2015). Involves 129 macroeconomic monthly time series data from 1959 to 2018. 8 Categories: Output and income, labor maket, housing, consumption, orders and inventories, money and credit, interest and exchange rate, prices in the stock market.

NBER Recession Dates:

• Labelling based on NBER dataset • 8 recession periods during the time period in consideration • 628 'normal' periods and '93' recession periods

Data Cleaning:

• Removal of variables with missing observation/ imputation of some sort • Add lags of all variables as additional features • Test stationarity of the time series • Standardize the dataset

Add lags of the variables as additional features:

• Add 3, 6, 9, 12, 18 month lags for each variable • Shift labels for 1 month ahead prediction • 699 observation points and 710 features

Stationarity:

• Augmented Dickey Fuller Test. Null hypothesis of ADFis that the time series is non stationary with the alternative that it is stationary • If p value > significance level, we cannot reject null hypothesis. Then take first order difference • adfuller function from statsmodels is used

Standardization:

• Standardization of feature vectors by removing mean and scaling to unit variance • StandardScaler from scikit-learn is used

Methodology:

• Perform feature selection to get the most important variables for the forecasts • Separate dataset into training and validation datasets. 1960 - 1996: Training, 1996 - 2018: Validation • Evaluate the performance of ML Algos on training set with Cross Validation • Select the best performing models based on average accuracy and std dev of the CV results. Logistic Regression chosen as benchmark • Make predictions on the validation dataset with selection models. Use GridSearchCV to find the best combination of hyperparameters. Evaluate the validation modela nd report accuracy metrics.

Cross Validation:

• K Fold CV used: ○ Train the model on (k-1) folds of the training data ○ The resulting model is validated on the remaining part of the data • 'TimeSeriesSplit' is CV technique for time series data. Use first k sets as training, (k+1) as test set

Evaluation Metric:

• ROC AUC Score