Utils for survival analysis of time series models

ts_survival_utils.py

def survival_mape_exp(df, win_size):
  '''Calculate expanding window mape for survival analysis.'''
  df['ape'] = np.abs((df['yhat'] - df['y'])/df['y'])*100
  df['exp_mape'] = df.ape.expanding(win_size).mean()
  df['survival_metric'] = df['exp_mape']
  return df


def survival_events(results):
  '''Generate events and censors for survival analysis.'''
  results = results[(results.index >= STUDY_START_WEEKS) & \
                    (results.index < STUDY_END_WEEKS)]
  thresh_res = results[results['survival_metric'] >= THRESH]
  
  # if some results are above threshold
  if len(thresh_res):

    # if all results are above thresh - this could be considered
    # as a case of left censorship - in our case we start evaluating
    # the model at STUDY_START_WEEKS and for this case it makes sense 
    # to force right censorship 
    if len(thresh_res) == len(results):
      thresh_week = STUDY_START_WEEKS
    else:
      thresh_week = thresh_res.index[0] + 1

    event = 1

  else:
    # event does not occur till the end of study
    thresh_week = STUDY_END_WEEKS
    event = 0

  return thresh_week, event