Honest by default
No black box. No hype.
We show you the data, the sample size, and the limits — and we hold every product to the standard we would want before trusting it ourselves. Every claim on this page is one a skeptic can check.
Everything we ship runs the same gauntlet first: an AI-driven research pipeline that reproduces, audits, and stress-tests a strategy until it survives — or gets rejected. It is what runs underneath NQ Edge+ and the Cone.
Independently verifiable on TradingView. Built to disprove its own findings.
Retail and semi-pro trading is saturated with strategies that look excellent in a backtest and fail in live markets. The causes are well known and rarely controlled for: lookahead bias, repainting, unrealistic fills, overfitting, and unverifiable claims — numbers no third party can reproduce. The result is an industry where a buyer can't tell a real edge from a lucky curve fit. We built SaneSignals on the inverse premise: distrust every result until it survives a gauntlet designed to destroy false positives.
Any one of these checks is standard. Holding research to all of them — and rejecting most candidates — is not. This is the core of what we do: a research operation engineered not to fool itself.
Each strategy is re-implemented from scratch by a second, independent process; the two must agree per-bar before any number is believed. A result that can't be reproduced is a lead, not a fact.
The engine is reconciled against TradingView's own bars and indicators within stated bounds — so a claim can be independently verified, not just asserted.
A documented protocol hunts the failure modes that make most backtests fake — lookahead boundaries, repainting signals, same-bar fill artifacts — and blocks any result that depends on them.
A reserved out-of-sample tail, metered so a strategy’s held-out data is spent at most once. An edge must hold its sign and rough magnitude on data it was never fit to.
Because we test many ideas, we control for multiple testing in our research process — a result must clear a bar that rises with the number of hypotheses tried. This separates a real edge from the best of a thousand coin flips.
We treat overfitting as the primary enemy — an edge tuned to noise is rejected, and we never optimize toward any single validation statistic.
Findings are run both recency-weighted and equal-weighted and flagged if their sign disagrees — catching effects that only exist in one regime.
Realistic commission and slippage are applied and stated openly. A “tradable” claim must survive the cost model.
Before any research task runs, an independent, context-blind reviewer audits the spec for ambiguity and wrong-question framing — catching bad questions before compute is spent.
Before any backtest, the underlying data and derived features are spot-checked against ground truth, so we're testing what we think we're testing.
Most candidates are rejected here — by design.
The pipeline is AI-compute-heavy by design. Large-language-model agents accelerate strategy reproduction and code translation, and operate the gauntlet itself — proposing candidate edges, blind-auditing their own research specs, independently reproducing results, and rejecting the majority under the statistical gates. Final ship decisions remain human-reviewed. The frontier we are building is a strategy-discovery engine that searches for genuine edges under the same anti-lookahead, anti-overfit constraints a human researcher would apply. Model inference runs on cloud infrastructure today; scaling that engine onto GPU-accelerated cloud instances is the primary use of growth capital.
Honest verification is only possible on well-reconciled data. We maintain a proprietary market-data layer: more than 5.4 million minute bars of NQ futures spanning roughly 16 years, reconstructed into continuous contracts with volume-dominance roll detection and reconciled against TradingView (within stated precision). Beyond NQ — the live product’s instrument, which is TradingView-reconciled — the same reconstruction pipeline extends to our other traded CME futures; those instruments are reconstructed but not yet TradingView-validated. The data layer is a defensible asset in its own right.
Most indicators draw a line or fire a signal. The Cone is a trade-context excursion model: for an active trade it characterizes how far price has historically travelled — in both directions — from comparable setups, conditioned on the current market state, to inform realistic profit-target and stop placement. It is not a single “containment %” readout; it describes the distribution of likely excursions, within the NQ Edge+ signal regime. Research in progress extends it with trend-conditioned directional skew — an early, out-of-sample-validated finding, not yet productized. We describe only what the data supports.
We show you the data, the sample size, and the limits — and we hold every product to the standard we would want before trusting it ourselves. Every claim on this page is one a skeptic can check.