Arima_functions.py

# Make a function to find the MSE of a single ARIMA model 
def evaluate_arima_model(data, arima_order):
    # Needs to be an integer because it is later used as an index.
    split=int(len(data) * 0.8) 
    train, test = data[0:split], data[split:len(data)]
    past=[x for x in train]
    # make predictions
    predictions = list()
    for i in range(len(test)):#timestep-wise comparison between test data and one-step prediction ARIMA model. 
        model = ARIMA(past, order=arima_order)
        model_fit = model.fit(disp=0)
        future = model_fit.forecast()[0]
        predictions.append(future)
        past.append(test[i])
    # calculate out of sample error
    error = mean_squared_error(test, predictions)
    return error

# Make a function to evaluate different ARIMA models with several different p, d, and q values.
def evaluate_models(dataset, p_values, d_values, q_values):
    best_score, best_cfg = float("inf"), None
    for p in p_values:
        for d in d_values:
            for q in q_values:
                order = (p,d,q)
                try:
                    mse = evaluate_arima_model(dataset, order)
                    if mse < best_score:
                        best_score, best_cfg = mse, order
                    print('ARIMA%s MSE=%.3f' % (order,mse))
                except:
                    continue
    return print('Best ARIMA%s MSE=%.3f' % (best_cfg, best_score))

   # Now, we choose a couple of values to try for each parameter.
p_values = [x for x in range(0, 4)]
d_values = [x for x in range(0, 1)]
q_values = [x for x in range(15, 20)]

# Finally, we can find the optimum ARIMA model for our data.
# Nb. this can take a while...!
import warnings
warnings.filterwarnings("ignore")
y_log = np.log(y)
evaluate_models(y_log, p_values, d_values, q_values)