Cognivec · Research · FX Results

FX Results Dossier

Headline results of the markets research line: geometric signals extracted from news text, tested against currency pairs. Results and conclusions are public; the constructions behind them are not — methods, features, and protocols are shared under NDA only.

Validation standard for everything below. All signals are causal (computed strictly before the market session they predict), evaluated walk-forward out-of-sample, net of transaction costs where marked "net", and checked against placebo machinery (shuffled calendars, random-anchor controls, permutation nulls). Pre-registration is used for confirmatory runs. These are research backtests, not live track records — a free 8-week daily forward-test is our standing offer to any fund that wants to judge for itself.

Signals that survived

Topological hole-direction → GBPUSD

net Sharpe +1.07 · 10-year walk-forward · p < 0.007 vs placebo

The flagship: the orientation of topological holes in the daily news embedding cloud predicts next-day GBPUSD direction. The hole's poles are nameable in plain words. The effect is specific — it does not transfer to other pairs, which is itself informative.

Curvature bridge-vector (independent second signal)

net Sharpe +0.72 / +0.53 across splits · p = 0.005 · corr 0.32 with the flagship

A geometrically distinct signal from the curvature of semantic bridges between topic clusters. Low correlation with the hole signal; the two-signal ensemble runs near +1.3.

Hard-pairs portfolio

net Sharpe +1.15 · p = 0.02 · positive in 4 of 5 out-of-sample years

Three signals with pairwise correlations between −0.16 and +0.15, combined on deliberately difficult currency pairs.

Semantic novelty ensemble

solo +0.79 · three-pair ensemble +1.97

How new today's news-meaning is relative to its own past predicts moves across three pairs; diversification across pairs triples the solo result.

Two-country six-level structure → cross rates

out-of-sample gross +2.54, net +1.00

Signals built from the joint structure of UK and US news at six timescales. Gross-to-net decay here is the honest headline: turnover eats most of the edge, and we say so.

Constant-axis signal (frozen years in advance)

honest out-of-sample net +1.47

An axis fixed on pre-2025 data only, then run forward untouched — the strictest test we know how to run against ourselves.

Hourly news geometry → intraday

net Sharpe +0.87 (hourly, K=168) · separate spread-filtered line +2.08 net over 7 years

The daily results replicate at hourly resolution with a different construction; a cost-aware execution filter is what makes the difference between paper edge and net edge.

Cross-signal entanglement

z = 4.08 · Sharpe +1.23

A dependence measure between news-geometry series and price — the only survivor of a nine-round adversarial search that killed 45 sibling hypotheses (see the graveyard).

The graveyard

We publish what died. Over a hundred hypotheses were tested and killed by the same machinery that validated the survivors — which is exactly why the survivors mean something. Highlights:

45 directional hypotheses, one autonomous nine-round search — all null. One survivor from an adjacent family (above).
Directional semantic axes — every train-selected axis collapsed out-of-sample. The close-twin control test proved decisive and is now standard in our protocol.
Cumulative-sum correlations with price — textbook spurious regression; killed by out-of-sample and block-bootstrap controls. Twice.
News → direction in high-volatility hours — null, despite a volatility forecast with IC 0.69 from the same features. Volatility is predictable; its sign is not.
Chaos-theory forecasting of news flow — Lyapunov exponent ≈ 0 across three countries: the news tape is stochastic, not chaotic. Full report.
Candle-as-token language models — a session-level candidate died on pre-registration.
Regime-gating by agreement — apparent gains were selection variance; random-gating placebo is now mandatory in our protocol.
An epoch effect confessed — one hole-based signal (causal gross SR 1.29, placebo 0) weakened on replication into the past: a real structure, but epoch-dependent. We report it as such, not as alpha.

What is deliberately not on this page

Feature constructions, anchor identities, axis definitions, window and refit parameters, selection procedures, and execution rules. The results above are verifiable in a forward-test without any of that — which is the evaluation we recommend. Methods open under NDA: [email protected].