Measuring the Accuracy and Predictive Coding at LegalTech NYC

February 10th, 2014

During LegalTech, there were probably half a dozen panels on E-Discovery with a focus on predictive coding. One of the more interesting panels offered a predictive coding case study, featuring Dr. David Lewis. With David’s permission, I post these tweets of the slides–though I would stress that these findings are preliminary and subject to change.

The first slide compares the precision rate for old-fashioned linear review, and for predictive coding. The former is at 76%, and the latter is at 87%. In other words, using predictive coding is more likely to identify the correct documents, and less likely to identify the wrong documents.

Second, predictive coding yielded superior results to even manual review, which is much more time-intensive.

The precision for identifying privileged documents was also much higher.

And predictive coding was better at identify the “hot doc” (that is the one document you do not want to produce).

In an earlier presentation, Judge Peck offered a helpful summary of predicting coding caselaw.

But there are still many pending questions.