Error: Initial guess outside of bounds

[capelastegui]

We are using the following custom model:

model_step_lockdown = forecast_models.get_model_step_date('2020-03-16')
forecast_models.model_linear+model_step_lockdown

On the initial value check, we get an out of bounds errors:

AssertionError: Initial guess outside of bounds: 
(((decay+step_lockdown)+(season_wday+season_month))+calendar_uk) 
- [-0.00114933         nan  0.          0.00120371  0.00236946  0.66702322
  0.01967633  0.02398207  0.24821339  0.95640447  0.59944057  0.8343352
 -0.07392551 -0.84970678  0.95439321  0.05401732 -0.03123115 -0.39606425
  0.60877698  0.06738449  0.35155036  0.58593382  0.          0.        ], 
[[-inf -inf   0. -inf -inf -inf -inf -inf -inf -inf -inf -inf -inf -inf
  -inf -inf -inf -inf -inf -inf -inf -inf -inf -inf]
 [ inf   0.  inf  inf  inf  inf  inf  inf  inf  inf  inf  inf  inf  inf
   inf  inf  inf  inf  inf  inf  inf  inf  inf  inf]]

Most likely the error is due to the second guessed parameter of the model being NA for some reason. The first 4 parameters are from model_decay, so let's focus on those:

Estimate:
[-0.00114933         nan  0.          0.00120371 ]

Bounds:
[[-inf -inf   0. -inf], [ inf   0.  inf  inf ]]

Relevant code for model_decay:

# - Exponential decay model: math::  Y = A * e^(B*(x-C)) + D
def _f_model_decay(a_x, a_date, params, is_mult=False, **kwargs):
    (A, B, C, D) = params
    y = A * np.exp(B * (a_x + C)) + D
    return y


def f_init_params_decay(a_x=None, a_y=None, a_date=None, is_mult=False):
    if a_y is None:
        return np.array([0, 0, 0, 0])
    A = a_y[0] - a_y[-1]
    B = np.log(np.min(a_y) / np.max(a_y)) / (len(a_y) - 1)
    if B > 0 or B == -np.inf:
        B = -0.5
    C = 0.
    D = a_y[-1]
    return np.array([A, B, C, D])

[capelastegui]

We can reproduce the problematic set of guessed parameters with an input array containing negative values:

forecast_models.f_init_params_decay(a_y=[1, -1, 0])

# Out: array([ 1., nan,  0.,  0.])

[capelastegui]

Solution: We need to check for non-negative input data when validating input for a model_decay