tblume1992 / ConformalImpact

Causal Impact but with MFLES and conformal prediction intervals

Conformal Impact

Take Causal Impact and replace the Bayesian Structural Time Series Model with MFLES and the Basyesian posterior with Conformal Prediction Intervals.

Quick Examnple an comparison to Causal Impact

intervention_effect = 400
np.random.seed(42)
series = np.random.random((130, 1)) * 400
x_series = series * .4 + np.random.random((130, 1)) * 50 + 1000
trend = (np.arange(1, 131)).reshape((-1, 1))
series += 10 * trend
series[-30:] = series[-30:] + intervention_effect

data = pd.DataFrame(np.column_stack([series, x_series]), columns=['y', 'x1'])

import matplotlib.pyplot as plt

plt.plot(series)
plt.plot(x_series)
plt.show()


from ConformalImpact.Model import CI


conformal_impact = CI(opt_size=20,
                      opt_steps=10,
                      opt_step_size=3)
impact_df = conformal_impact.fit(data,
                              n_windows=30,
                              intervention_index=100,
                              seasonal_period=None)

conformal_impact.summary()
conformal_impact.plot()





from causalimpact import CausalImpact

impact = CausalImpact(data, [0, 99], [100, 130])
impact.run()
impact.plot()
print(impact.summary())
output = impact.inferences
np.mean(output['point_effect'].values[-30:])