When Machine Learning Is the Wrong Briefing Slide

Machine learningGovernanceEthics

Machine learning can sharpen prediction; it does not automatically clarify mechanisms. In cross-over labs between ML and econometrics tracks, we ask participants to write two slides: one predictive, one causal. Mixing them on the same slide without a bridge is the most common failure mode we see. We teach leakage checks ruthlessly because they are boring until they are not. Temporal splits matter for macro and finance examples common in Thai desks. Random k-fold can look great on paper and be nonsense in production when calendars matter. Interpretability tools help, but they are not transcripts of reality. SHAP values summarize model behaviour conditional on the data generating process you fed the model. We practice saying that aloud in meetings until it feels natural. The closing exercise is a “decision memo” that uses ML outputs as inputs to a separate human judgment about interventions. If participants cannot articulate what would change their mind, we send them back to the checklist. It is slower than a flashy deck, and it ages better.