Adversarial ML & Robustness in Trading — Final Exam

This is the graded finale for Adversarial Machine Learning & Model Robustness in Trading, and it runs across everything the course built. You started by adopting the security mindset: a market is not a benign sampler of i.i.d. draws — it is full of counterparties who profit when your model is wrong, so you must reason in threat models (who attacks, what they know, what they can change) and accept that robust is not the same as accurate. Reflexivity and the Adaptive Markets view mean the very act of trading on an edge can destroy it. From there you met adversarial examples — the smallest input nudge that flips a decision — formalized by FGSM, where the perturbation is $\delta = \epsilon\,\mathrm{sign}(\nabla_x L)$, and its iterated cousin PGD; you saw that a linear model’s worst-case loss shifts by $\epsilon\lVert w\rVert_1$, so high-norm weights are fragile, and that $\epsilon$ is best read as an attacker’s trading budget. You studied training-time attacks — data poisoning, backdoors that fire on a trigger, and spoofing/layering that is simultaneously an adversarial input and a poisoning of your features — plus confidentiality attacks: model extraction (cloning your model from its fills) and membership inference (detecting your training set, amplified by overfitting). You built the language of distribution shift — covariate, label, and concept drift — and learned why live strategies decay (overfitting, alpha crowding, reflexivity), monitored with PSI and CUSUM. Finally you stocked the robustness toolkit — adversarial training as a $\min_\theta \max_\delta$ game, distributionally robust optimization over an ambiguity set, regularization that bounds sensitivity, ensembling that cuts variance toward $\sigma^2/k$, and conformal prediction whose coverage guarantee leans on exchangeability — and you learned to stress-test honestly: worst-case rather than average, alert to gradient masking, red-teamed, and reported on a deflated scorecard. No hints are shown, each answer locks the moment you submit, and your score stays hidden until the very end.

Course Recap