Talks & Activities
2026
-
Presenting on the Panyu campus Invited Talk September 2, 2026Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted ScoresCollege of Cyber Security, Jinan University · Guangzhou, ChinaI delivered an invited research talk in the 2026 Academic Lecture Series of the College of Cyber Security at Jinan University, presenting my work on temporal leakage in LLM backtesting — how pretrained models silently "read the future," how to measure the contamination, and how to eliminate it — to faculty and graduate students, followed by an extended Q&A.
More about this talk
The talk treated temporal leakage at two levels. At the level of an individual prediction, I presented tools that measure how much leaked evidence actually drives a forecast, filter leaked claims at inference time, and train the model to reason only from pre-cutoff information. At the level of the backtest score, I showed that the usual pre/post contamination check is uninformative — legitimate recency effects reproduce the leakage signature — and that one external reference, such as a known training cutoff or a matched clean control, restores measurement and yields a leakage-adjusted score. The 90-minute session ran in person on the Panyu campus with a parallel online audience, and closed with an extended Q&A on evaluation practices for deployed forecasting systems.
