regime-change

Summary

I am forecasting US recessions based on macroeconomic indicators from 1984-2020

Results

Logistic Regression model trained on 1984-2009 data forecasted 2020 recession 1 month in advance of the official data from FRED

The Model

I have used Logistic Regression model with l2 penalty trained on 1984-2009 data forecasting 2020 recession. I have also assigned a higher weight to recession class to account for inbalanced data.
Splitting data into train and test set (data up to the end of 2009):

df_train = df[:322].drop(['Date', 'Regime'], axis=1)
df_train_targ = df.iloc[:322, 1]
df_test = df[323:].drop(['Date', 'Regime'], axis=1)
df_test_targ = df.iloc[323:, 1]

Fitting the model into Train set and forecasting Test set:

model=LogisticRegression(C=0.029, penalty='l2', class_weight={0: 0.15, 1: 0.85}) # high weight to recession 
LR=model.fit(df_train,df_train_targ)
training_predictions=LR.predict(df_train)
predictions=LR.predict(df_test)
prob_predictions = LR.predict_proba(df_test)
print(f'Accuracy Score: {accuracy_score(df_test_targ,predictions)}')
print(f'Confusion Matrix: \n{confusion_matrix(df_test_targ, predictions)}')
print(f'Area Under Curve: {roc_auc_score(df_test_targ, predictions)}')
print(f'Recall score: {recall_score(df_test_targ,predictions)}')
print(f'Precision score: {precision_score(df_test_targ,predictions)}')

Accuracy Score: 0.98
Confusion Matrix:
[[119 --- 1]
[1 --- 8]]
Area Under Curve: 0.94

Macroeconomic indicators (Features)

After running the model for the first time and extracting features importance I selected the following indicators picking the ones with low correlation:

Bucket	Indicator	Details
Bond Market and Monetary Policy	T1YFFM	1-Year Treasury C Minus FEDFUNDS
Bond Market and Monetary Policy	TERMSPREAD	Term Spread (10-year T-bill minus 3-month T-bill)
Bond Market and Monetary Policy	T10YFFM	10-Year Treasury C Minus FEDFUNDS
Bond Market and Monetary Policy	T5YFFM	5-Year Treasury C Minus FEDFUNDS
Real Estate Market	HOUST	Housing Starts: Total New Privately Owned
Employment	AWOTMAN	Avg Weekly Overtime Hours : Manufacturing
Employment	ICSA	Initial Claims, Monthly
Employment	UNRATE	Unemployment Rate
Economic growth	SP500	Index Growth
Foreign Exchange	EXUSUK	U.S. / U.K. Foreign Exchange Rate
Employment	PAYEMS	All Employees, Total Nonfarm
Employment	USCONS	All Employees, Construction

The positive scores indicate a feature that predicts class 1, whereas the negative scores indicate a feature that predicts class 0
Classes:
1 - Recession
0 - Normal Regime

# get importance
importance = LR.coef_[0]
feature_imp = pd.DataFrame({'Feature': df_new.drop(['Date', 'Regime'], axis=1).columns.to_list(), 'Importance':importance})
most_imp = feature_imp[(feature_imp['Importance']>0.025)|(feature_imp['Importance']<-0.1)]
most_imp.set_index('Feature').plot(kind='bar', figsize=(10,5))
plt.title('Feature Importance')
plt.ylabel('Importance')

Feature extraction and resampling

recession = mitk.recession_nber() #monthly
T10YFFM_d = mitk.tenyear_cm_minus_fedfunds() #monthly
TB3SMFFM_d = mitk.three_month_tbill_minus_fedfunds() #monthly
PAYEMS_d = mitk.all_employee() #monthly
JTSJOL_d = mitk.job_open() #monthly
T5YFFM_d = mitk.five_year_cm_minus_fedfund() #monthly
T1YFFM_d = mitk.one_year_cm_minus_fedfund() #monthly
EXUSUK_d = mitk.us_uk_fx() #monthly
HOUSTW_d = mitk.housing_stats_west_census() #monthly
TB3MS_d = mitk.three_month_trs_secondary() #monthly
USCONS_d = mitk.employee_constr() #monthly
AWOTMAN_d = mitk.overtime_manuf() #monthly
HOUST_d = mitk.housing_owened() #monthly
DMANEMP_d = mitk.employee_durable() #monthly
SRVPRD_d = mitk.employee_service() #monthly
ICSA_d = mitk.initial_claims() #daily
ICSA_d = ICSA_d.resample('M').mean() #resampling daily to monthly
UNRATE_d = mitk.unemployment_rate_us() #monthly
sp500 = fltk.hist_returns(['^GSPC']) #daily
sp500_m = (1+sp500).resample('M').prod()-1 #resampling daily to monthly
sp500_m = sp500_m.rename(columns={'^GSPC': 'SP500_M'}) #changing column name
termspread = mitk.spread_tenyr_threemon_tr() #daily
termspread = termspread.resample('M').mean() #resampling daily to monthly
gdp = mitk.gdp_us() #quaterly
gdp = gdp.resample('M').mean() #resampling quaterly to monthly
gdp = gdp.interpolate(method='nearest') #interpolating gaps
gdp = gdp.pct_change() #percent change

Data is pulled from FRED using their API. All functions are located in macro_indicators_toolkit.py file. Sample function:

def three_month_trs_secondary():
    """
    3-Month Treasury Bill: Secondary Market Rate, Monthly
    """
    threems = fred.get_series('TB3MS')
    threems.dropna(inplace=True)
    threems = threems.rename('TB3MS')
    return threems

I have added 3, 6, 9, 12 and 18 months lags to each feature to enhance the model prediction capabilities:

# Add lags
for col in bigmacro.drop(['USREC'], axis=1):
    for n in [3,6,9,12,18]:
        bigmacro['{} {}M lag'.format(col, n)] = bigmacro[col].shift(n).ffill().values

# 1 month ahead prediction
bigmacro["USREC"]=bigmacro["USREC"].shift(-1)
# dropping columns with missing values
bigmacro=bigmacro.dropna(axis=0)

I have standartized the data to adjust for different scales:

# Standardize
from sklearn.preprocessing import StandardScaler
features=bigmacro.drop(['USREC'],axis=1)
col_names=features.columns

scaler=StandardScaler()
scaler.fit(features)
standardized_features=scaler.transform(features)
standardized_features.shape
df=pd.DataFrame(data=standardized_features,columns=col_names)
df.insert(loc=0,column="Date", value=bigmacro.index)
df.insert(loc=1,column='Regime', value=bigmacro['USREC'].values)
df.head()