The crypto market is flooded with AI signal services promising superior returns through machine learning, neural networks, and large language models. Monthly subscriptions range from 50 to 500 dollars for signals that claim to predict price direction using artificial intelligence. The implicit promise is that AI sees patterns that humans and traditional indicators miss.
We use AI in our trading pipeline. We have measured its contribution precisely. The honest answer is nuanced: AI adds real value in specific, bounded roles, but it is not the source of our edge and it is not worth paying a premium for as a standalone signal.
What Our AI Actually Does
Our platform uses Claude (Anthropic's LLM) for four tasks: sentiment analysis (scoring news headlines from negative 1.0 to positive 1.0), signal enrichment (adjusting strategy confidence by plus or minus 0.2), regime narration (contextualizing quantitative regime labels), and post-trade analysis (reviewing completed trades for quality and learning).
The signal enrichment is the only component that directly affects trading decisions, and its impact is bounded. The maximum adjustment is 0.2 confidence points. A strategy generating a signal at confidence 0.6 can be adjusted to anywhere between 0.4 and 0.8. AI cannot reverse the signal direction, cannot create signals that do not exist, and cannot bypass any risk check.
This means AI contributes at most a 20 percent adjustment to a decision that was already made by a rules-based strategy. The other 80 percent or more of the trading decision comes from RSI, MACD, Bollinger Bands, and the quantitative risk framework.
Measuring the AI Contribution
We track whether AI adjustments improve outcomes by recording the original confidence, the AI adjustment, and the eventual trade PnL. Over our paper trading period, the data shows that AI adjustments are directionally correct approximately 55 to 60 percent of the time. This is better than random (50 percent) but far from the 70 to 80 percent accuracy that AI signal services imply.
The 55 to 60 percent accuracy on a bounded 0.2 adjustment translates to a modest improvement in overall strategy performance. The exact magnitude is difficult to isolate because it interacts with position sizing (confidence affects size) and risk gates (confidence affects whether marginal signals pass thresholds). Our estimate is that AI enrichment improves portfolio Sharpe by 0.1 to 0.3 points. Real but not transformative.
What AI Signal Services Actually Provide
Most AI signal services fall into one of three categories. The first is a rules-based system with an AI-generated description. The strategy uses RSI, moving averages, or similar indicators, but the marketing calls it AI because the description was written by ChatGPT. The AI contribution to the actual trading decision is zero.
The second is a genuine ML model (typically a gradient-boosted tree or neural network) trained on historical price data to predict direction. These models can produce positive backtest results through overfitting to historical patterns. The critical question is whether the model has been validated out-of-sample across multiple market regimes. Most have not.
The third is an LLM-based analysis where ChatGPT or a similar model is asked to analyze market conditions and produce a buy or sell recommendation. This is the most marketed and least reliable category. LLMs produce plausible-sounding analysis regardless of whether the prediction is accurate. The analysis quality is determined by the training data, not by any special market insight.
Why We Do Not Sell AI Signals
We could package our AI enrichment as a signal service. We choose not to because the value proposition is dishonest. Our strategies produce Sharpe ratios from 1.7 to 19.0. The AI contribution is approximately 0.1 to 0.3 Sharpe points of that. Selling the AI component as the source of the edge would misrepresent where the value actually comes from.
The value comes from the rules-based strategies (Bollinger Bands, RSI, MACD), the rigorous validation pipeline (five-period regime testing), the risk management framework (eight layers of defense), and the operational infrastructure (45 bots running 24/7 with crash recovery). AI enhances the system. It does not drive it.
When AI Signals Are Worth Paying For
An AI signal service is worth considering if it meets all of the following criteria. The signals have been backtested with realistic fees and slippage. The backtest has been validated across multiple market regimes (not just the recent bull market). The provider publishes drawdown statistics alongside returns. The methodology is transparent enough to evaluate. The signal includes risk parameters (stop-loss, position size), not just direction.
If any of these criteria are missing, the signal service is selling hope rather than edge. The vast majority of AI signal services fail on criterion two (no multi-regime validation) and criterion four (black-box methodology).
The Bottom Line
AI is a useful tool in a trading system. It provides contextual enrichment, sentiment analysis, and anomaly classification that improve decision quality at the margin. It is not a source of standalone trading edge. The strategies that produce our highest Sharpe ratios use AI for at most a 0.2-point confidence adjustment on decisions that were already made by validated rules-based logic.
If someone is selling you AI trading signals without showing multi-regime validated backtest results with realistic execution costs, they are selling the AI hype, not the AI edge. The edge, when it exists, comes from the quantitative framework underneath. AI makes it slightly better. The framework makes it work.