Starting from first principles: Why #AI Trading should not predict price movements

1. Breaking down AI: What are large models really doing?
Many people consider large models as "thinking black boxes," but from a computational perspective, they do only one thing:
Calculate similarity in high-dimensional vector space.
All inputs (#token , images, time series) are mapped into vectors;
All "understanding" and "reasoning" are essentially vector dot products + nonlinear transformations.
The core of Transformer is not "intelligence," but:
Compress inputs into vectors
Compute similarities between vectors
Aggregate information based on similarity
Output a probability distribution
This reveals a key fact:
LLMs' ability is pattern recognition, not prediction.
In language scenarios, patterns are stable enough to seem like prediction;
but in financial markets, this approach breaks down completely.
2. Why "using AI to predict rise or fall" almost certainly fails
Because financial markets and language have three fundamental differences.
1️⃣ Very low signal-to-noise ratio
Natural language has strong statistical regularities;
Market short-term prices ≈ random walk, with up/down roughly 50:50.
Models often learn only:
Mean reversion
Smoothing of noise
Not the extreme volatility and risk events you truly care about.
2️⃣ Non-stationarity
Language semantics remain stable over decades;
Market structures constantly evolve, effective patterns in 2021 may be invalid in 2024.
3️⃣ Strong adversariality
Language has no counterpart;
Markets are zero-sum games, any identified pattern will be quickly arbitraged away.
The conclusion is straightforward:
AI is not suitable for predicting "the next K-line will go up or down."
But this doesn't mean AI is useless in trading—just that the question is wrong.
3. Reframing the question: Recognize Regime, not predict direction
The truly high signal-to-noise problem is:
What is the current market state?
Markets are not continuous random walks but switch between different Regimes:
Low volatility oscillation
High volatility oscillation
Unilateral trend
Liquidity crisis
Regimes are persistent and have time scales much longer than a single K-line.
This makes them much more stable than "predicting rise or fall."
Extensive research shows that the alpha from Regime-based strategies mainly comes from avoiding bad states, not capturing good ones.
4. Correct way to use AI: Market State Embedding
Transfer the computational paradigm of LLMs, but change the goal.
Not:
Market data → AI → Up or down
But:
Market data → Vector representation → Similarity → Regime judgment
Core idea:
Use an encoder to compress multi-dimensional market states into embeddings
Vector distances represent "market state similarity"
Is the current market similar to which historical phase?
This step:
Does not assume distributions
Does not set artificial thresholds
Naturally supports high-dimensional, multi-factor, nonlinear data
Outputs are not "buy/sell signals," but market state awareness.
5. The true value of Regimes: Strategy routing and risk control
Regimes do not directly generate profits, but they determine whether you'll lose big.
Typical mappings:
Oscillation → Grid trading
Trend → Trend following
High volatility → Reduce positions
Liquidity crisis → Hold cash
The focus is not on earning more, but on:
Not doing the wrong thing in the wrong market state.
6. NoFx: The infrastructure layer of AI Trading, not an "AI predictor"
NoFx's positioning is clear:
Not to have LLM predict markets, but to provide AI with a "safe decision-making trading system."
Why is the bottleneck of AI Trading not the model?
Because a functioning system requires:
Stable, unified data
Low-latency, controllable execution
Strict, configurable risk management
Complete, auditable logs
99% of "AI trading products" lack these.
7. What NoFx is doing (simplified)
1️⃣ Data layer
Unified encryption of all core market data:
Prices, volumes, OI, funding rates, liquidations, capital flows, order books, volatility, technical indicators.
Heterogeneous data → unified interface.
2️⃣ Execution layer
Hide exchange differences, unify order placement, position, leverage, stop-loss logic.
Strategies don't need to care "which exchange."
3️⃣ Decision layer
AI does not predict rise or fall, but:
Structured market analysis
Asset screening and ranking
Position and risk assessment
Entry/exit condition judgment
Outputs are auditable structured decisions, not just "buy."
4️⃣ Regime integration
Regime recognition as a system module:
First determine market state
Then decide strategy type and risk level
AI's aggressiveness automatically adjusts with Regime
8. Why insist on transparency and open source?
In finance, black boxes lack trust.
NoFx principles:
Every AI decision has full context
Every trade can be traced back to "why"
Risk management, position, leverage all configurable
Code is auditable and deployable by oneself
This is what infrastructure should look like.
9. Democratization: AI Trading should not be exclusive to institutions
Reality:
Quant trading has extremely high barriers
Ordinary traders are systematically excluded
NoFx's simple goal:
Enable someone with no coding knowledge to set up their own AI trading strategy in 5 minutes, and understand what it’s doing.
This is not about lowering professionalism, but encapsulating expertise into tools.
Like:
Excel for data analysis
Figma for design
NoFx aims to become the standard layer of AI Trading.
10. One sentence summary
The essence of AI is vector similarity, not "predictive ability."
High-value market questions are Regime recognition, not rise/fall prediction.
The real barriers are not models, but engineering, risk management, and interpretability.
The future of AI Trading is infrastructure, not hype.
If you're interested, I can write a dedicated next article:
👉 Why "AI just buy when told" is the most dangerous trading system design