AI trading signals: what they are actually worth

Artificial intelligence has changed many trades, and market analysis is no exception. Model-generated signals are no longer confined to trading floors: they are available to individuals, sometimes free. What remains is to understand how they are built, what you can reasonably ask of them, and what it would be unwise to expect.

What a trading signal is#

A signal is an indication to buy or sell an asset, accompanied by reasoning and usually by levels: a target, an invalidation threshold, a horizon. A signal produced by artificial intelligence comes from a model that combines many variables at once — price history, volumes, technical indicators, cross-market correlations, news — to form that indication.

The difference from human analysis is not the quality of the reasoning, it is the volume processed. An analyst follows a few dozen assets. A model sweeps several thousand continuously, without fatigue and without preference.

How a signal is built#

The chain has three stages. First collection: live prices, computed indicators (RSI, MACD, Bollinger bands, moving averages), news feeds, and cross-market correlations. Then analysis: the model looks for known configurations and assesses context. Finally formatting: a direction, a confidence level, explicit reasoning, numerical levels and a horizon.

The confidence level is the most widely misunderstood part. It does not state the probability that price will rise: it states how far the analysed elements converge. High confidence on a mistaken reading is still a mistaken reading.

The families of signals#

• Technical signals, drawn from charts and indicators alone.
• Fundamental signals, derived from company financial data.
• Sentiment signals, measuring the tone of news and publications.
• Correlation signals, exploiting the way an event in one market propagates to another — a rise in oil and airlines, for instance.

Can they be trusted?#

Not blindly, and the qualification matters. What these models do well: process large amounts of data quickly, spot configurations a human eye would miss, work without emotion and without interruption.

What they do badly: anticipate the unforeseeable, and respond to a situation without precedent. A model recognises what resembles what it has already seen; faced with something genuinely new it still produces an answer, with the same assurance. That is its main danger.

How TradeSynapse produces its signals#

TradeSynapse signals are generated by an Anthropic Claude model, from prices, computed indicators, news aggregated from more than 190 sources, and a set of 22 cross-market correlation configurations. Every signal is published with its reasoning: that is the useful part.

The built-in assistant runs on Claude Haiku for the entry tiers and on Claude Sonnet from the Pro tier upwards. Signals are capped at three per day on the free tier.

One point of transparency worth noting: past signal performance is published with its decomposition, including the share of signals that expire without reaching either their target or their invalidation threshold. That share is substantial, and a platform that does not publish it is giving you an artificially flattering success rate.

Using a signal properly#

1. Read the reasoning before the direction. If the reasoning does not hold, the direction does not matter.
2. Cross-check against another source, or at minimum against your own reading of the chart.
3. Size the risk yourself: never more than 1 to 2 % of capital on one trade, whatever confidence level is displayed.
4. Test in simulation for at least a month before following a signal with real money.
5. Keep a journal: it is the only way to know whether signals actually help you, rather than assuming they do.

What comes next#

Three developments are taking shape. Multimodal models, able to read a chart as an image alongside numerical data. Personalisation, where the same signal is framed differently according to the horizon and risk tolerance of whoever reads it. And explainability, which European regulators are already demanding: a signal whose reasoning cannot be reconstructed will be hard to offer to the public.

That last requirement points the right way. An explained signal is a signal you can dispute — which is what separates it from an oracle.