We can make our code from the The time-dependent temperature mean adjusted autoregressive model section more robust by doing the following:
Computing the median instead of the mean
avgs[i-1] = np.median(temp[indices])
Ignoring the outliers with a masked array
temp[:cutoff] = ma.masked_array(temp[:cutoff], temp[:cutoff] < (q1 - 1.5 * irq))
We get slightly different output with the modified code, with about 70 percent of the values predicted having an absolute error of less than 2 degrees Celsius:
AR params [ 0.95095073 -0.17373633] % delta less than 2 70.8567244325