LLMs trained on word sequences can also be pointed at price sequences. Hudson River Trading laid out the math at NeurIPS 2023: 3,000 stocks, 10 data points per stock per day, 252 trading days, 23,400 seconds per session yields roughly 177 billion market tokens per year. GPT-3 trained on 500 billion tokens. The data volume is comparable. The predictability is not.
The fundamental problem is adversarial signal decay. Language has grammar; no one actively obscures the next word. Financial markets have thousands of well-funded competitors erasing every exploitable pattern, producing what economist Lasse Pedersen calls an 'efficiently inefficient' system. Noise dominates. The data distribution also shifts constantly with regulatory changes, macro shocks, and irrational flows like the 2021 GameStop episode. Language evolves over decades. Market regimes change in weeks.
The more interesting arguments in this piece are structural, not predictive. The parallel between financial factor residualization and transformer residual networks is precise and worth reading slowly. So is the section on attention mechanisms mapping to multi-scale market phenomena, where earnings signals decay over months, momentum over days, and order book imbalance over seconds. The unsolved problem the author identifies: transformer models predict one period ahead, but profitable quant strategies require forecasting the full return term structure simultaneously. That gap has no clean solution yet.
[READ ORIGINAL →]