FINDEC · method and evidence
An agentic financial decision system, evaluated without lookahead.
FINDEC plans which evidence a question needs, gathers it from specialist agents, reconciles their disagreements, and weighs them by their measured track record. It is evaluated by a forward test whose predictions are sealed before their outcomes exist — so no result on this page could have been informed by the outcome it reports.
Two planes, separated on purpose
The separation is what makes the numbers interpretable.
Control plane
Decides what work to do. Language-model driven.
Not backtestable, and not asked to be: a model trained on data covering the test window recalls rather than forecasts. It is evaluated on task quality instead — planning accuracy, routing, adjudication.
Decision plane
Decides what the market will do. Numerical and deterministic.
Deterministic given its inputs, so its out-of-sample performance means what it appears to mean. Sentiment is scored locally by FinBERT, which is frozen and carries no knowledge of the evaluation window.
Why this matters. Put a current language model in the decision path and backtest it on 2018–2023 and it is not forecasting — it already knows what happened. Keeping the numerical claims free of it is the difference between a result and an artefact.
Weights that move
No constant anywhere in the fusion step.
Each agent's influence is a function of observable state: its self-reported confidence, its measured hit rate in the prevailing volatility regime, and the freshness of the data behind it. Self-reported confidence and earned reliability are kept in separate terms deliberately — otherwise a confident agent outvotes an accurate one.
A language model is never asked to emit a weight. A number produced that way would be unbacktestable, unstable between runs, and impossible to audit after the fact; a weight derived from measurable state is none of those and adapts just as much.
The forward test, live
Read directly from the sealed store. Whatever it currently shows is what it shows.
Loading live evaluation data…
Per-agent accuracy
Each agent judged on what it said, not on whether it was followed.
Intent taxonomy, grounded rather than invented
The Planner classifies a query into one of eleven intents, derived from Blankespoor et al.'s study of 29,242 real retail-investor questions rather than from intuition. On that paper's own published labels the Planner scores 0.923, against the 0.75 their trained SBERT classifier achieves on the same task.
This figure is not yet quotable. The prompt was revised four times against those same 26 questions, so it is tuning-set performance and carries the same optimism as a strategy tuned on its own backtest. A held-out set the prompt has never seen is required before the number means anything.
The same data also shows what FINDEC does not yet serve: explaining and screening are 75% of what investors actually ask, and market data outweighs news 31:1 as an information source.
What this does not yet show
Stated here rather than left to be discovered.
- No performance claim. The forward test has produced no scored outcomes yet. Nothing here says the agentic arm beats the numerical control.
- Risk-adjusted claims are out of reach. An earlier five-ticker study had 5.9% power to detect a +0.30 Sharpe difference. Breadth was widened to 40 names precisely because duration cannot fix that, and directional accuracy — not Sharpe — is what this design can support.
- Agent reliability is still a prior. Every weight currently uses 0.50 for measured reliability, because no agent has a track record yet.
- The Auditor is not yet a trustworthy instrument. It returned different verdicts on identical input across two runs; that variance needs explaining before any audit figure is reported.
- Fundamentals are point-in-time incapable. The source reports only current values, so that agent refuses historical dates outright rather than serving a figure that would be quiet lookahead.