Hi all! Would this be an appropriate implementation of the Simple Combination of Univariate Model (SCUM) if I have monthly data (all positive) and I want to generate 2 years worth of predictions?

# Monthly dataset:
df = df[['unique_id', 'ds', 'y']]

seasonality = 12 

models = [
    AutoETS(model = 'ZZZ', season_length = seasonality),
    DynamicOptimizedTheta(season_length = seasonality),
    AutoCES(season_length = seasonality),
    AutoARIMA(season_length = seasonality)
]

# Instantiate StatsForecast class
sf = StatsForecast( 
    df = d_new,
    models = models,
    freq = 'MS', 
    n_jobs = -1,
    fallback_model = SeasonalNaive(season_length = seasonality)
)

sf.fit()

d_sf = sf.predict(h=24)
model_cols = [c for c in d_sf.columns if c != 'ds']

d_sf['yhat'] = d_sf[model_cols].clip(0).median(axis=1, numeric_only=True)

Hello! I haven't run the code on my end, but looks like you're taking the median of different statistical models, which seems reasonable to me!

@DANIEL KIM I think it's pretty much 95% there. Missing some extra logic they use at least in the paper for some sample size and seasonality based edge cases: https://eprints.lancs.ac.uk/id/eprint/131943/1/IJF_2019_SCUM_post_print_.pdf

Sounds good, thank you for your feedback! I'll check out automfles as well

Haven't heard of SCUM method before, thanks for the links!

It’s really nice to see how it performs!