Predicting Options

The computer science department has provided an efficient alternative to preregistration

It’s the year 2004, and though we have yet to see flying cars and Rosie the Robots in our kitchens, at least we can now better predict course enrollments. Thanks to an innovative algorithm designed by students in Computer Science (CS) 96, “System Design Projects,” the College will be able to prevent the misallocation of 1,200 teaching fellows (TFs) in future terms.

The catalyst to this project was last year’s preregistration debacle. For years, course prediction methodology has not been accurate enough to allow departments to reliably hire and train TFs far enough in advance —leaving TFs in a precarious job market and students in precariously-led sections. The solution offered by the administration last spring (withdrawn after wide criticism) was forcing students to choose their courses a semester in advance. But that system did not account for the inevitable enrollment swings that would have resulted from the liberal add/drop period it proposed in place of shopping, and the plan would have worked to the detriment of student flexibility. Where preregistration failed, this new initiative prevails.

The algorithm, dubbed a “machine learning system” by McKay Professor of Computer Science Stuart M. Shieber ’81, who teaches CS 96, incorporates more than a decade of data—including historic enrollment, course type, department, CUE Guide ratings and time slot. The gains in efficiency of a standardized mathematical model for predicting course enrollment will benefit the entire Harvard community.

The administration should be commended for supporting this course prediction project and looking beyond the short-sighted preregistration proposal. The problems that preregistration ostensibly sought to improve are valid and serious and the University is right to search for better methods of predicting enrollment.


Specifically, Harvard should not hesitate to employ non-binding surveys asking students in which courses they intend to enroll. The machine learning system will undoubtedly struggle to predict enrollment in new courses and these surveys could be used as a supplement.

Moreover, the TF hiring process could be much improved if the University centralized the practice. A standardized online system would allow professors and TFs across departments to connect more easily. TFs would be able to rank their preferences for multiple positions and an automated system would match them appropriately.


The preregistration debate of last spring was an extremely healthy one, the fruits of which are only now being realized. Legitimate problems are being addressed—and in far more amenable ways. As course enrollment prediction becomes more accurate and the TF hiring process becomes more streamlined, we all stand to gain from an enhanced educational experience.