Machine learning approach for small samples ARMA models identification

Abstract EN

This paper proposes an effective machine learning
approach to identify small samples data generated from
autoregressive moving-average ARMA(p, q) models.

The
theoretical and practical aspects of the proposed approach are
introduced, and its validity was evaluated by the ratio of
correct identification(CIR).

For evaluating the validity of the proposed machine
learning approach, a simulation study was achieved.

192000
small samples were generated from ARMA(p, q) models with
different sample sizes(10, 20, 30) and different parameters sets
through the stationarity and invertibility regions.

The ratio of
the correct identification is calculated and used for evaluating
the proposed approach.

The average of CIR for all samples
was 99.3% which shows a good performance for the
proposed approach.

The results also showed that the
automatic ARMA identification Is less sensitive to small
samples additionally, The proposed approach is quicker,
automatic and more accurate alternative.

A Python program
is written for doing automatic Identification using a machine
learning attached in the appendix.