•
Feng et al. add intentional perturbations during training so stock-movement models generalise better under noisy price features—a robustness habit you can borrow for NLP signals too.
Feng, Chen, He, Ding, Sun, and Chua propose adversarial training for stock up/down movement models. Because price-based features are noisy and change continuously, a network trained only on clean snapshots can overfit. Their Adv-ALSTM idea: train the model to stay correct under small, intentional perturbations of latent features. Relative accuracy gains are reported versus a strong baseline on public stock datasets. For after-work readers, the transferable lesson is robustness under noise — useful when you later score news text as well. This Hub note is educational reading for after-work Python practice. It is not investment advice, not a trading signal, and not an offer to buy or sell any security.
Evening ML notebooks often treat yesterday’s close, returns, and indicators as fixed numbers. Markets do not. Tiny measurement differences, microstructure noise, and regime shifts mean the same “feature vector” is never perfectly repeatable. A model that memorises one clean path can look great in-sample and fragile out-of-sample. The problem is not “more layers”; it is training that ignores stochasticity.
The paper’s Adv-ALSTM (Adversarial Attentive LSTM) adds adversarial perturbations to high-level representations during training, so the classifier must work on both clean and slightly attacked examples. That is the same family of ideas as adversarial training in computer vision, adapted to sequential stock features with attention.
Empirically, the authors report average relative accuracy improvements on the order of a few percent versus a prior state-of-the-art attentive model on real stock datasets (see the paper for exact tables). Code for their setup has been released publicly (Adv-ALSTM on GitHub) for people who want to read an implementation — still for study, not for a trading desk.
You can borrow the habit even if you never train an LSTM tonight: when you build a news or price feature, ask “what happens if I nudge this input a little?” That question alone upgrades after-work experiments.
Authors: Fuli Feng et al. (IJCAI 2019; arXiv:1810.09936). Landing: arXiv abs.
One evening exercise (educational only):
sklearn).Optional later: skim the Adv-ALSTM repo structure; do not port TensorFlow 1.x on night one unless you enjoy dependency archaeology.
When you want structured evening lessons on Python itself — variables, loops, functions, files, and first libraries — start with Introduction to Python Programming. Quant and macro courses are not the CTA here; build the language first.