Uber is in the prediction business, but it wants to get better. So the company's artificial intelligence team has proposed a new algorithm that it says improves its forecasting of demand in extreme situations like bad weather and major holidays.
Uber researchers are applying a machine-learning architecture called "long short-term memory" (LSTM) to predict demand for drivers during "holidays, concerts, inclement weather, and sporting events." In a new blog post, they said that this new approach improved the accuracy of their existing forecasting algorithm between 2% and 18%.
Why it matters: Better AI forecasting models could impact fields as diverse as medicine and earthquake prediction. The latest craze among hedge funders is using LSTM architecture to predict asset price changes.
How it works: Traditional statistical modeling requires humans to decide the relevance of historical data to the given forecast. In the LSTM architecture, a predictive algorithm improves itself by finding data relationships otherwise unrecognized by humans. In Uber's case, these data are the number of completed rides, app views, current riders, outside temperature, and wind speed.