Research
Research Is the Product, Not the Marketing
Financial machine learning is easy to fool and easy to fool yourself with. Most of what looks like edge in a backtest is leakage, luck, or an evaluation rule that quietly changed. Our research process exists to rule that out before anything is trusted.
Why financial ML needs strong anti-leakage controls
A model that has, even indirectly, seen information from the future will look better than it is. Because market data is sequential and autocorrelated, leakage is easy to introduce by accident — through a shared scaler, a feature built with a forward-looking window, or a label that peeks past the decision point. We treat leakage prevention as a first-class engineering requirement, not an afterthought.
Chronological testing
Every split, fit and evaluation step follows time order. Nothing from a later period is used to inform an earlier decision, including preprocessing and hyperparameter selection.
Research and Production acceptance are separate
A method performing well in research is a research result, not a Production claim. Promotion from research to any higher-trust status is a distinct, explicit decision — never inferred automatically from a good metric.
Negative results
Most tested methods do not qualify. We record why: which check failed, which assumption broke, which evidence was insufficient. A negative result that is understood is more valuable than a positive result that is not.
Model and method comparison
Candidate models and methods are compared under the same frozen rules and the same data boundaries, so that differences reflect the method — not an inconsistent evaluation setup.
Independent audits
Material research conclusions are reviewed by someone who did not build the thing being reviewed, checking hashes, splits, predictions and metrics rather than trusting a self-report.
Robustness and causal reconstruction
Accepted methods are stress-tested against level and scale shifts, noise, missing data and temporal extrapolation, and features are reconstructed to reflect only causally available information at each point in time.
Why failed experiments improve the system
A rejected method still teaches us something about the data, the market regime, or a blind spot in our own process. Publishing that lesson is part of Build in Public — and part of why our accepted results can be trusted more, not less.
The Evidence Ladder
A result does not become a claim by being interesting. It climbs this ladder — one rung at a time, in order — or it does not get promoted.
Hypothesis
A candidate idea or method is formalized under an explicit, versioned contract before any data is touched.
Backtest
The idea is tested against historical data under frozen, leakage-safe evaluation rules defined before the result is seen.
Leakage-safe walk-forward validation
Chronological, causally-reconstructed validation confirms the result is not an artifact of the evaluation setup itself.
Independent audit
A reviewer who did not build the method checks hashes, splits, predictions and metrics — not a self-report.
Final approval
A distinct, explicit decision accepts, rejects, or sends the result back for more work. Nothing promotes itself.
Live-paper observation
Only after acceptance is a method observed against live markets — still virtual, still publicly reported.
Our Position
Tantoryn AI does not try to reinvent machine learning. We use strong existing research and proven methods, then apply Tantoryn-specific causal, leakage-safe validation and compatibility testing before anything is trusted with real market decisions.
Research discipline aside, see the live MAIN/SHADOW reports →