Define what the system actually does
Course context: This lesson belongs to a practical beginner curriculum for Pakistan.
| Claim | Evidence required |
|---|---|
| “AI signal” | Input data, target variable, update time and exact output |
| “High accuracy” | Class balance, test dates, false-positive rate and untouched sample |
| “Adaptive” | When retraining occurs and how old/new versions are separated |
| “Live performance” | Broker statements including spread, commission, slippage and subscription cost |
A valid research split
Keep training, validation and final test periods separate by time. Use walk-forward testing so every prediction uses only information available at that moment. Reusing the final test set for model selection turns it into training data and invalidates the reported result.
Stress the result before demo
- Increase spread and commission assumptions.
- Add execution delay and adverse slippage.
- Remove the best few trades.
- Test a volatility regime not dominant in training.
- Shift signal thresholds slightly.
- Compare the model with a simple non-AI baseline.
If a small change destroys the result, the model is not ready for a demo pilot.
Demo deployment controls
- Set a hard maximum lot size and total exposure.
- Reject orders with missing or stale prices.
- Use an idempotent order ID to prevent duplicates.
- Stop after a daily cash-loss threshold.
- Log signal, model version, quote, order response and error.
- Require manual reactivation after the kill switch.
Fraud checks for AI products
Reject guaranteed returns, a 100% win rate, secret results without broker evidence, payments to an individual wallet, pressure to recruit others, or claims that “no trading knowledge is required”. The CFTC warns that fraudsters use AI language to promote bots and signal services with unrealistic or guaranteed performance.
Primary sources
CFTC: AI will not turn bots into money machines · Exness API capabilities and eligibility
Questions from a first-time learner
Does machine learning remove market risk?
No. It can only model relationships in available data; regimes, prices and execution can change.
Is accuracy the best metric?
Not alone. Evaluate precision, payoff, costs, drawdown and errors by market regime.
Can I test an AI strategy on demo?
Yes, but first lock the model version and add risk and logging controls.
What is the most important live control?
A hard exposure limit and a tested kill switch that stops new orders.