Neural network-based time series forecasting of student enrollment with exponential smoothing baseline and statistical analysis of performance.

Abstract

The sustainability of educational institutions generally depends largely on strategic planning, both in terms of optimal allocation of resources/manpower and budgeting for financial aids/scholarships to incoming students. Hence, forecasting of student enrollment plays a vital role in making crucial decisions based on previous time-bound records. This work demonstrates the power of neural network-based time series forecast over a traditional time series model and recommends the better network architecture between deep and shallow neural networks based on 25-year historical records of student enrollment in Programming Fundamentals from 1995 – 2020 at Kansas State University, Manhattan Campus. The study reveals that Vanilla Long Short-Term Memory (LSTM) model performs better than the deep neural network with Root Mean Square Errors (RMSE) of 0.11 and 0.24 respectively – both of which produced better results than the Single Exponential Smoothing baseline having a RMSE of 0.27. The study also carries out a statistical analysis of 5-year student performance based on weekly Labs, Projects and Mid-Terms using Analysis of Variance (ANOVA). The result shows the existence of differences in the yearly average performance of students. Post Hoc Tukey’s pairwise multiple comparison tests reveals consistency in performance up to the period of the semester where possible dropouts would have occurred. Students’ delay in tackling challenging projects also accounts for the significant differences in the mean scores.