Chainfeeds Summary:
The bottleneck in AI Trading is not the model, but the engineering.
Article source:
https://x.com/Web3Tinkle/status/2007960269235126350
Article Author:
Tinkle
Opinion:
Tinkle: After understanding the computational process of large models, a core question arises: what exactly do large models do? On the surface, it seems to be predicting the next word; for example, given today's weather, the model outputs the highest probability of "good," seemingly predicting the future. However, if we break down its internal mechanism, we find that it actually performs three steps: First, it encodes the current context into a high-dimensional vector; second, it searches the parameter space for the pattern that best matches this vector; and finally, it outputs the probability distribution corresponding to this pattern. Therefore, large models are not inferring unknown events, but rather identifying which known pattern in the training data is most similar to the current input. Prediction is inferring the uncertainty of the future; recognition is matching historical patterns. The reason why large models excel in the language domain is that natural language itself has extremely strong and stable statistical regularities—after certain word sequences, the distribution of subsequent content is highly concentrated. When you ask ChatGPT a question, it is not "thinking about answers," but matching the language pattern most similar to your input and outputting the most common response form under that pattern. On the surface, it appears to be prediction, but at its core, it is large-scale pattern recognition. Applying this pattern recognition logic to financial markets seems intuitive: encode market data into vectors, retrieve the most similar states in history, and output the probability distribution of future trends. However, a fundamental difference exists: the statistical regularities of financial markets are far weaker than those of natural language. In language, the uncertainty of the next word is often very low, while in the market, the rise and fall of the next candlestick is highly close to randomness. Numerous studies have shown that models on financial time series often only learn weak patterns such as mean reversion, and are almost powerless against truly important extreme fluctuations. More importantly, markets are highly non-stationary and antagonistic: effective patterns quickly become invalid as participant structure, regulatory environment, and capital behavior change, and any pattern that can be systematically exploited will be wiped out by arbitrage forces. Therefore, directly predicting rises and falls using large models is destined to fail. But this does not mean that pattern recognition has no value in finance. The key is: don't ask where prices will go, but ask what state the market is currently in. Compared to predicting single-point rises and falls, identifying the market's current state has a higher signal-to-noise ratio. Markets oscillate between various states, including low-volatility oscillations, high-volatility oscillations, upward trends, downward trends, and liquidity crises. These states often exhibit persistent and identifiable structural characteristics. Based on this idea, a market state embedding can be constructed: high-dimensional heterogeneous market data is compressed into low-dimensional vectors, making similar periods close to each other in the vector space. Through comparative learning, clustering, or similarity retrieval, it's possible to determine which historical state the current market is closer to, and to make strategy selections and risk control accordingly. Its value lies not in capturing every market movement, but in proactively reducing positions or exiting during high-risk periods to avoid systemic drawdowns. NoFx's positioning revolves around this concept, building the infrastructure layer for AI Trading: avoiding AI's mystical market predictions, and instead helping traders make more robust choices in complex and volatile markets through transparent and interpretable period identification and structured decision-making.
Content source 
