From Analysis to Prediction: How Does AI Generate Tradable Signals?

I. Prediction Problems vs. Trading Problems Are Not Equivalent

In machine learning, tasks are usually defined as predicting the next n periods’ price movement, returns, or volatility.

But in trading, the problem structure is more complex, with at least four decision layers:

• Whether to trade;
• Trade direction and position size;
• Holding duration and exit conditions;
• When to reduce frequency or pause the strategy.

This means even with high prediction accuracy, stable returns may not materialize.

A typical scenario: the model correctly identifies most small movements but repeatedly fails during rare large swings, with tail losses erasing accumulated profits.

Therefore, the first conclusion of Lesson 3 is: signal value comes from “executability,” not a single prediction metric.

II. Three Main Signal Generation Paths

Rule-based Models

Rule-based models are built on clear logic, such as “breakout + volume confirmation + risk filter.”

They feature strong interpretability, fast deployment, and low debugging costs.

Their limitation is weak ability to capture complex nonlinear relationships and require manual rule reconstruction when market structures change.

ML-based Models

Machine learning approaches learn nonlinear relationships via multiple factors, commonly seen in tree models, time-series models, and probabilistic scoring models.

The advantage is handling higher-dimensional inputs and uncovering combinations beyond manual rules.

Risks include overfitting, reduced interpretability, and higher strategy maintenance thresholds.

Hybrid Frameworks

Hybrid frameworks typically use “rule constraints + model ranking” or “model timing + rule execution.”

This path is more common in practice because it balances flexibility and robustness:

• Rules handle boundaries and baseline;
• Models handle probability judgments and priority sorting.

In most intermediate trading systems, hybrid frameworks are more sustainable than pure rule or pure model approaches.

III. Signal Target Design: Beyond Direction—Include Risk and State

Defining labels as “future up/down” is often too crude. Practical designs usually include three target layers:

1. Direction: probability of rising or falling in the next n periods;
2. Magnitude: expected return range or potential drawdown range;
3. State: whether the market is trending, ranging, squeezed, or experiencing liquidity contraction.

These layers have distinct roles:

• State layer decides if the strategy is activated;
• Direction layer decides position bias;
• Magnitude and risk layer decide position weighting and stop-loss strength.

This upgrades a single prediction task to a layered decision system, significantly reducing the odds of “correct direction but losing trades.”

IV. From Model Scores to Trading Actions: Thresholds and Confidence Management

Model outputs are usually probabilities or scores—not direct buy/sell commands.

Signal implementation hinges on threshold management and tiered execution:

• Low confidence: no trade or observation only;
• Medium confidence: small position trial;
• High confidence: increase weighting but still subject to risk caps.

The core idea is “signal layering,” not “equal weighting.”

If all signals are executed equally, it leads to noisy trades, fee erosion, and excessive turnover.

Mature systems focus on net signal quality—reducing low-quality trades and boosting per-trade effectiveness.

V. Signal Evaluation Requires More Than Accuracy

Tradable signal evaluation should cover five dimensions:

1. Prediction: accuracy, recall, error distribution, probability calibration;
2. Trading: win rate, profit/loss ratio, turnover rate, holding period structure;
3. Returns: cumulative profit, max drawdown, return/drawdown ratio;
4. Execution: fee sensitivity, slippage sensitivity, capacity limits;
5. Stability: performance consistency across different market states.

If a signal excels only in prediction metrics but fails trading or execution dimensions, it lacks live trading value.

Many “great in backtest but fail live” strategies stem not from the model itself but from missing the signal-to-trade mapping chain.

VI. Common Failure Mechanisms: Model Failure Often Means Market Structure Shift

Live trading failures typically stem from three types of shifts:

• State shift: trending markets turn to ranging markets, original momentum signals lose validity;
• Crowding shift: more participants adopt similar strategies, compressing marginal returns;
• Cost shift: after volatility drops, original trading frequency results in sharply higher cost ratios.

Thus, signal systems must have failure monitoring mechanisms such as:

• Is hit rate and profit/loss ratio declining together?
• Is signal distribution abnormally clustered?
• Is trade quality and slippage deteriorating persistently?

When thresholds are triggered, reduce frequency/leverage or pause to avoid compounding losses during failure windows.

VII. Unique Signals in Crypto Markets: Price, Position, On-chain Three-Dimensional Interaction

Compared to traditional markets, crypto’s advantage is simultaneous observation of price action, derivatives positions, and on-chain fund flows.

For example, common composite signals include:

• Price breaking key zones;
• Rising open interest and heating funding rates;
• Strengthening net stablecoin inflows.

This structure may indicate trend reinforcement—or fragile surges after high leverage crowding.

Therefore, directional signals must pair with risk filters:

• Has funding rate entered overheated zone?
• Is liquidation cluster too close to current price?
• Can order book depth absorb impact?

Only after passing risk filters does the signal gain higher execution value.

VIII. From Single Model to Signal Factory: Executable Iteration Path

The course recommends an iterative “small system first” approach rather than aiming for complex architecture immediately. Executable steps:

1. Select a single task (e.g., 4-hour directional probability);
2. Establish an interpretable baseline (rules or lightweight model);
3. Complete threshold grading and trade mapping;
4. Perform out-of-sample and rolling validation;
5. Observe live small-position “signal-execution-result” linkage;
6. Gradually expand into multi-signal integration (direction + risk + state).

Core advantages are diagnosability, replayability, iterability—consistently reducing uncertainty from black-box decisions.

IX. Lesson Summary

This lesson focused on “How AI generates tradable signals.” Key conclusions include:

• Signal generation isn’t just a prediction task—it’s a decision system engineering challenge;
• Rule-based, machine learning, and hybrid frameworks each have boundaries; hybrid paths are more practical for implementation;
• Tradable signals must meet prediction, trading, execution, and stability requirements simultaneously;
• In crypto markets, signals should be based on price, position, on-chain three-dimensional interaction—with matching risk filters.

Key takeaway: Predictability does not equal tradability; signal value is determined by execution quality and risk control together.

The next lesson will move into the next step of a complete closed loop: strategy automation—from backtesting to live trading—focusing on how to engineer signal systems for continuous operation.