mutable struct MultiplexModel <: SupervisedModel
    models::Vector{SupervisedModel}
    best_model::Int
end

# when instantiating, specify that there's no best_model yet (see predict)
function MultiplexModel(models::Vector{SupervisedModel}, best_model=0)
    MultiplexModel(models, best_model)
end

function fit!(mm::MultiplexModel, X, y)

    X_A, y_A, X_B, y_B = train_test_split(X, y)

    quality = [0 for m in mm.models]

    @parallel for (i, model) in enumerate(mm.models)
        fit!(model, X_A, y_A)
        quality[i] = metric_of_choice(y_B, predict(model, X_B))
    end

    mm.best_model = findmax(quality)[2]

    fit!(model, X, y)

    mm
end

function predict(mm:MultiplexModel, X)
    # could be something prettier here
    @assert mm.best_model > 0 "mm not fitted yet"
    predict(mm.models[mm.best_model], X)
end