Which possible future deserves attention?
A forecasting model can draw a plausible path without producing a useful decision. I wanted a system that would retain its predictions, wait for the actual outcome, and learn from the difference. The challenge was not only generating candidates; it was selecting among them without quietly learning from the future.
Traid builds on the open-source Kronos candlestick foundation model. My contribution is the terminal, data flow, candidate selection, calibration, replay, persistence, and evaluation around that model: the system that turns generated paths into a research question I can test.
A loop that keeps its own receipts.
Each forecast generates 12 candidate paths across the next 12 candles. Eight markets and three timeframes have independent selectors, supported by feed adapters, a FastAPI service, persisted history, and a live relay to the interface. A resolved forecast only becomes training evidence after all 12 future candles have matured.
The selector starts with a simpler Ridge model. A gradient-boosted alternative is promoted only when chronological validation supports it and enough labels and independent windows exist. Research on richer context and more complex models keeps the same discipline: training-only transformations, conservative overrides, and a protected fresh holdout.
The selector improved. That is a specific result.
The documented three-run batch evaluated 768 cases per run: 2,304 cases and 27,648 candidate-path labels in total. Candidate-quality mean absolute error fell from 13.947 to 11.785, a 15.5% decrease. Selector regret fell from 13.735 to 12.631, about 8.0%. These are research scoring measures, not portfolio returns.
By the third run, 14 of the 24 selectors qualified for the gradient-boosted model; ten kept Ridge. Keeping the simpler model where the evidence was weaker was part of the design. Each promotion had to earn its place through chronological validation.
Candidate-quality MAE across the three research runs
Lower is better. A research scoring measure from the documented three-run batch; not a measure of investment return.
Forecast improvement is not proof of profitable trading.
A later cost-aware paper diagnostic, documented in the September 30 product audit, did not establish a profitable strategy. It recorded a profit factor of 0.801 over 844 historical simulated trades after modeled costs. That dataset had already informed research, so it was not fresh out-of-sample proof either.
Those findings shape the next stage: protected holdout evaluation, prospective paper qualification, and explicit execution limits. The case study is about building a research system that can expose its own weaknesses. It does not claim live trading qualification, realized profits, or a validated investment strategy.
About this case study +
Repository research papers and reports: research/README, ACCURACY_V4, ARCHITECTURE, V14_RESEARCH, PAPER_TRADING, and the September 30, 2026 product audit. Later cost-aware findings are presented alongside earlier selection results.
Researched and updated October 3, 2026.