Machine learning, deep learning and language models work at every stage of Radar61’s forecasting, alongside statistics and econometrics. Every model is benchmarked, monitored and explained, and its forecasts are scored like any other.
Stage 1 of 7 · Information collection
What AI does here
Held to accountEvery input keeps its source and timestamp.
Statements, filings, news and reports become structured inputs: who and where, what happened, how it reads, how it reaches markets, and what it resembles. Every extraction links back to its source.
Example source · shipping report
Attacks on vessels in the Red SeaRegion have pushed more container shipsSector around the Cape of Good HopeRoute, adding about ten days to Asia–Europe voyagesTrade lane. Freight rates rose for a third week, and European manufacturersSector warned of parts shortages.
Structured output · entities
No single model wins everywhere. The method follows the problem, the data, what can be measured and what it costs to run, and every candidate has a simpler baseline to beat.
Will gold be higher or lower in four hours?
A complex model can memorise the past and still fail on the future. Walk-forward testing trains on one window and tests on the next, again and again, and models that don’t beat the baseline there don’t ship.
Error relative to the baseline · lower is better · example
On the data it was trained on, the deep network looks best by far.
See what moved a forecast and test it: change the assumptions and the forecast, its range and its attribution respond. Try it.
Assumptions · example EURUSD forecast, next 24 hours
Deployed model · rolling calibration error · example
Markets change, and a model that worked last quarter can quietly stop working. Calibration, inputs and predictions are monitored, and drift is flagged and reviewed before it reaches the forecasts.
Ask a question and get a research note, not a chat reply: the answer, the forecasts it rests on, the evidence, and how uncertain it is.
Answer
A hotter print would push rate expectations up. That usually firms the dollar and weighs on gold, though gold can hold up if it reads as a policy mistake.
Uncertainty
Wide: the reaction depends on the size of the surprise and the Fed's response.
Forecasts it rests on
Evidence
Chosen for the problem, measured on the record, and combined in ensembles where that helps.
AI works alongside statistics, econometrics and fundamental research. The method follows the problem, the data and what can be measured.
Every model is compared with naive and statistical baselines. Added complexity has to earn its place on data it has not seen.
Training and testing use only the information that was available at the time, so backtests don't flatter the model.
Calibration, inputs and predictions are monitored. When the market changes and a model drifts, it is reviewed.
Explanations show what moved a forecast. They are grounded in the data and never claim more causality than the evidence supports.
Every forecast carries its range and probability, and every outcome is scored, so AI is judged by the record like everything else.
Forecasts across every market, with the range stated and every outcome scored.
