Issues when merging eval tables - CH02

[DiogoParreira]

I have noticed 2 potential errors in the notebook CH02

1. You do the below in your nb
   zero_shot_preds = zero_shot_preds.rename(columns={"NBEATS": "NBEATS_pretrained"})
   finetuned_preds = finetuned_preds.rename(columns={"NBEATS": "NBEATS_finetuned"})
   trained_preds = trained_preds.rename(columns={"NBEATS": "NBEATS_trained"})

test_df = pd.merge(test_df, zero_shot_preds, 'left', 'ds')
test_df = pd.merge(test_df, finetuned_preds, 'left', 'ds')
test_df = pd.merge(test_df, trained_preds, 'left', 'ds')

test_df.head()

when doing this, this will throw an error because all 3 df have unique_id as a column. I'd do something like this

keys = ["unique_id", "ds"]

test_df = test_df.merge(zero_shot_preds[keys + ["NBEATS_pretrained"]], on=keys, how="left")
test_df = test_df.merge(finetuned_preds[keys + ["NBEATS_finetuned"]], on=keys, how="left")
test_df = test_df.merge(trained_preds[keys + ["NBEATS_trained"]], on=keys, how="left")

2. In the section - Forecasting another frequency
   you do

pretrained_model = NeuralForecast.load(path='./model')

d_zero_shot_preds = pretrained_model.predict(d_input_df)

and you can see in the head of the df that the predictions are on a monthly basis

unique_id
1981-12-31	15.576023
1982-01-31	16.021452
1982-02-28	16.849257
1982-03-31	17.739635
1982-04-30	17.888920

but then you merge it with predictions of daily timesteps that you did here

models = [NBEATS(input_size=2*horizon, h=horizon, max_steps=500)]

nf = NeuralForecast(models=models, freq='D')
nf.fit(df=d_input_df)

d_trained_preds = nf.predict()

d_zero_shot_preds = d_zero_shot_preds.rename(columns={"NBEATS": "NBEATS_zero_shot"})
d_zero_shot_preds = d_zero_shot_preds.reset_index(drop=True)
d_trained_preds = d_trained_preds.rename(columns={"NBEATS": "NBEATS_trained"})

d_test_df = pd.merge(d_test_df, d_trained_preds, 'left', 'ds')
d_test_df = pd.concat([d_test_df, d_zero_shot_preds['NBEATS_zero_shot']], axis=1)

It shouldnt work and it will throw an error when trying to run the code.