Three very different things get called AI
| Type | What it does | Where it is strong | Where it breaks |
|---|---|---|---|
| Rule-based robot (EA) | Follows fixed rules written by a person | Discipline, speed, never tired or scared | Rules can stop working when markets change |
| Machine learning model | Learns patterns from past data | Finding structure in large datasets | Overfits: learns the past's noise, fails on new data |
| LLM agent (language model) | Reads and writes text, can call tools | Summaries, explanations, monitoring news, boring chores | Weak at precise numbers, can sound confident while wrong |
What AI genuinely does well in trading
- Explaining. Turning a price move or a trade into plain language, quickly and at any hour.
- Organising information. Summarising news, tracking many markets, flagging unusual conditions.
- Checking and monitoring. Watching a system for errors that a person would miss at 3 a.m.
- Testing ideas faster. Writing and running experiments, which is how we ran the test below.
What it cannot do (yet)
- See the future. Markets are noisy and competitive. A model with a small edge is common, a model with a large stable edge is extremely rare.
- Guarantee profit. Any seller who says otherwise is selling something else.
- Reason reliably with exact numbers. Language models can invent a figure and state it with confidence, which is why numbers should come from code, not from the model.
- Replace risk management. The most advanced model still needs position sizing and a stop.
A test we ran on ourselves
To check how much a careful, data-driven approach could find, we built a forecaster for gold. For each day it finds past days with a similar volatility, trend and position in the recent range, looks at what happened next, and turns that into probabilities. Then we tested it on 3,595 past days using only information that was available on the morning of each day, so it could not peek.
| What we measured | Result | What it means |
|---|---|---|
| Probability of up / flat / down | About 36% to 45% for 'up', almost always | Close to the long-run average every day |
| Direction skill, next day | 0.16% better than the average | A tiny edge, not a tradable one |
| Likely range (50% band) | Held 49% of the time | Honest: it said 50% and got about 50% |
| Likely range (80% band) | Held 78% of the time | Honest again |
How we use AI on this site
Our trading is done by a rule-based EA. The AI never decides a trade: it explains what the EA did, answers questions, and helps write articles around numbers our own code computes. Every figure on this site comes from our data, and a post is not published if its numbers do not match the source.
Ten questions to ask any AI trading product
- Is there a verified live track record on a third-party tracker, on a real account, with a track record longer than a few months?
- Are the back-tests on real tick data with realistic spreads and slippage?
- Was it tested on data it never saw (out-of-sample or forward-tested)?
- What is the maximum drawdown in testing and in live trading, and is it on the same account type you will use?
- How many trades is the result based on? Under a hundred proves nothing.
- What happens when the market changes? Does it stop, adapt, or keep trading blindly?
- Who holds the money? It should always stay in your own broker account.
- What does the AI actually do? Predict, execute, or just explain?
- Is anything guaranteed? If yes, walk away.
- Can you watch it live? A seller who hides the live account is telling you something.
Frequently asked questions
Can AI predict the gold price?
Not reliably. In our walk-forward test over 3,595 days a careful model had almost no edge on direction (0.16% better than a simple average) but gave honest ranges.
What is the difference between an EA and an AI trading bot?
An EA follows fixed rules written by a person. An AI bot may use a machine-learning model or a language model. Many products use the label AI for what is really a rule-based EA.
Is an AI trading agent safe?
Only with strict limits on position size, a stop on every trade, and funds that stay in your own account. Language-model agents can state wrong numbers confidently, so they should not set risk.
How does GoldTriEngine use AI?
The EA trades by fixed rules. AI explains the trades and helps write content around numbers our own software computes. It never places a trade.
Sources and method
- Our walk-forward test of the gold outlook engine: 3,595 days, k=250 similar days, daily bars from 2010; accuracy measured with the Brier score against a long-run average. Code and scorecard are public on this site.
- Brier, G. W., Verification of forecasts expressed in terms of probability, Monthly Weather Review, 1950