Research guide
AI crypto predictions: how to evaluate their limits
Scenarios, out-of-sample tests and costs: what to check before trusting an algorithmic result.
A scenario is not a promise
A target depends on the model’s inputs and assumptions. Changing volatility, horizon or trend can materially change the result. An AI-written explanation does not validate those calculations. A number labelled confidence can only be interpreted as a calibrated probability when supporting evaluation exists; it should not be confused with the chance of making money after fees.
What to require from a backtest
Separate data used to fit a model from data used to evaluate it. Check whether each signal was available at the simulated time, whether failed assets are included and whether fees, spread and slippage are charged. Comparing with holding the asset or a simple rule helps establish whether complexity adds value. A good result in one bull market does not demonstrate robustness.
Accuracy, loss and costs are different
Getting direction right six times out of ten does not imply profitability: a few large losses can exceed many small gains. Examine drawdown, outcome distribution, sample size and sensitivity to costs. Evaluating a forecast requires its issue time, horizon, target and observed-price source, preserved without rewriting them after the outcome becomes known.
External data and operational safety
CryptoOráculo tools use external data and automated analysis; delays and interruptions can occur. This guide is not an audited performance record or a trade-execution service. If evaluating an external bot, start without withdrawal permissions and check its behaviour without real funds. Never enter a recovery phrase to view a prediction.
Before making a decision
- Inputs and assumptions identified, including missing values.
- Out-of-sample evaluation and comparison with a simple baseline.
- Losses and costs disclosed, not only selected winning trades.