•
Sirignano & Cont (2018) train one LSTM on pooled NASDAQ books and beat per-stock models—even on names never seen in training. Universality + path dependence. Not investment advice.
Takeaway. One universal deep net trained on pooled NASDAQ books predicts the next mid-price move better than stock-specific linear or deep models—and it still works on names the trainer never saw. History of the book matters; ad-hoc vol/spread normalisations did not help.
Universal features of price formation in financial markets: perspectives from Deep Learning (Justin Sirignano, Rama Cont, 2018; arXiv:1803.06917). Pair with Sirignano’s LOB net.
Fit for learners: Shows pooled-stock deep models beat per-ticker fits — mindset shift for building one Python model across names.
This Hub note is a plain-language intro. Not investment advice.
Industry default: estimate one model per ticker on a short recent window because “markets are non-stationary.” That fragments data so deep nets never see enough scenarios. Microstructure theory often writes a generic map from supply/demand to price—but practice refuses to pool.
A scientific result about F—not a day-trading signal.
This is permission to concatenate panels in pandas instead of a folder of 500 tickers. Transfer to a new listing is the same idea as a function that does not hard-code the symbol.
Tiny evening exercise: two DataFrames a, b with columns ret; pd.concat([a.assign(sym='A'), b.assign(sym='B')]) and groupby-mean. Write why pooling can regularise a hungry net. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. Sirignano–Cont is academic microstructure research—not a recommendation to trade a universal LSTM.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—loops over files become one concatenated table; then models can share weights.