Signals feed
Direction, time horizon, calibrated confidence — and one click to the full audit chain behind it. Honest pagination; no infinite scroll of filler.
Causal market intelligence
See the whole chain,
See the whole chain,
not just the headline.
Priors reads the news you care about and reasons outward from it — one economic step at a time, without ever being shown a ticker — until a chain lands on something you can actually trade. Then it backtests every link. Only what survives becomes a signal.
Every branch is grounded in dated evidence, and no relationship becomes a signal until it survives event studies against historical prices under false-discovery-rate control. Plausible is not good enough.
Causal chain
Illustrative example
Strait of Hormuz disruption Event
Reuters · shipping & energy desk
Brent crude Commodity
Mechanism — seaborne crude supply shock, freight re-routing
US ethanol margins Factor
Mechanism — blending economics track the crude complex
Corn futures Mapped exposure
Mechanism — ethanol is the marginal buyer of US corn
Link survived validation
p 0.004
n 41
FDR-controlled
The move you missed was three hops away.
The direct relationships are already priced in by the time you read the story. The ones that actually pay are the indirect ones — and they are exactly the ones a person cannot trace by hand.
01
Too many degrees of separation
A shipping-lane story and a corn contract are four steps apart across three asset classes. Nobody is holding that graph in their head while the market moves.
02
Plausible stories are cheap
Ask any model for a causal narrative and it will give you a beautiful one. A story that sounds right and a relationship that actually held historically are different things.
03
You can't audit a black box
A score with no chain behind it is a coin flip with extra confidence. If you can't see the mechanism and the evidence, you can't decide whether to believe it.
How it works
Eight stages. The model never sees a ticker until stage six.
The pipeline runs continuously over the sources you configure. The expensive parts are gated by the cheap parts, the persuasive part is gated by the statistical part — and the reasoning happens before any asset is on the table, so a chain can never be built backwards from a symbol someone wanted to justify.
01
Ingest
Your RSS and Atom feeds are polled on a schedule, deduplicated and clustered, with per-source health tracking so a dead feed never silently starves the graph.
RSS/Atom · source packs
02
Gate
A keyword pass then local embeddings score each item against the interests configured on the platform. Most news never reaches a model at all — which is what keeps this affordable.
local embeddings · relevance gate
03
Extract
Survivors go to a model that pulls out what actually happened — the event, its actors, its timing — grounded in spans of the article rather than summarised from memory.
typed extraction · cited spans
04
Reason outward
The event is expanded into what it concretely affects next, then what that affects — with a mechanism, direction, lag, observable and falsifier on every hop. The model doing it has never heard of your watchlist. Each chain is then reviewed adversarially for what it conveniently left out.
catalog-blind · adversarially reviewed
05
Ground
Each branch is checked against dated, real reporting drawn from beyond your own feeds — counterevidence included. Where no evidence exists, that is recorded as a coverage gap instead of papered over.
independent evidence · dated citations
06
Map
Only now do assets enter. Chains that reached a terminal, measurable economic effect get matched to investable vehicles from a shared catalog — with the exposure channel named, including the negative ones.
terminal steps only · shared catalog
07
Validate
Every mapped link is backtested against historical price bars with event studies, and the whole family of tests is FDR-controlled. Most hypotheses die here. That is the point.
statsmodels · event study · FDR
08
Signal
Survivors become a directional, horizon-bounded signal with calibrated confidence — and a full audit chain: every hop, its evidence, and the article it came from.
calibrated · auditable
Why it works
Two tools that are each useless alone.
This is the whole idea, and it is the reason the product is called Priors: a hypothesis is a prior, the backtest is the evidence, and only the posterior ships.
Prior The language model proposes
An LLM carries the world knowledge to guess that ethanol blending sits between a shipping lane and a grain contract. No statistical method will ever invent that link, because it isn't in the price data.
What it cannot do: tell you whether the link is real. Ask it for a confidence and it will produce a number that means nothing. Show it a ticker first and it will cheerfully build a chain that arrives there — which is exactly why it is never shown one.
Posterior The statistics decide
Event studies over historical bars measure whether the proposed relationship actually held — with false-discovery-rate control, so testing thousands of hypotheses doesn't manufacture winners.
What it cannot do: imagine the hypothesis in the first place. It can only judge what it is handed.
Each covers the other's blind spot. The model only ever proposes; the data disposes. Nothing that fails validation is allowed to become a signal — plausibility alone is never enough.
The product
What you actually get after you sign in.
One dashboard: the feed you read daily, the research behind it, and the machinery underneath both whenever you want to interrogate a claim. Everything carries a label saying which it is.
Deepest research paths
The run's chains ranked by how far they actually reach and how well they're supported, with near-duplicates of the same story kept from crowding out the rest. Depth is not confidence, and the ranking never pretends otherwise.
Graph explorer
Walk the graph: recent events, the assets they drive, and every forward-reachable hop. Trace a chain node by node, mechanism by mechanism — with what's been validated and what's still a proposal marked apart, and each hop carrying when it should first show up versus when it fully plays out.
Price charts with event windows
See the bars the validation actually ran on, with the event window overlaid — so you can check the claim against the tape yourself.
Research syntheses
Related reporting reconciled into one cross-document picture: where the sources agree, where they contradict each other, and which claims are reported fact versus first-principles inference. Read-only research — it never moves a signal.
Your watchlist & sources
You choose the vehicles you're watching and the feeds you read — or install a reviewed source pack and see every publisher in it first. Your watchlist decides what's worth researching and what you're shown; it is never handed to the model that does the reasoning.
Runs you can watch
Trigger analysis on demand or leave it live, and watch it move stage by stage. Every run reports its yield in plain counts — articles researched, publishers, mechanism steps, economic layers, grounded steps, coverage gaps — so a thin run looks thin.
Coverage gaps
The mechanisms the research reached for and could not evidence — kept, with what would confirm them and what would kill them. A question nobody has reported on yet is worth more to you than a silent hole in the graph.
Built to be doubted
Every claim comes with the receipt.
A signals product is only worth anything if you can attack it. These are the invariants the system holds itself to.
Validation is a gate, not a garnish
An unvalidated edge cannot appear inside a signal. There is no "high confidence because the model said so" path through the system. Research that hasn't cleared that bar is still shown to you — labelled as a proposal, and never given a confidence number.
Full audit chain on every signal
Each hop, the evidence behind it, and the article it was extracted from — down to the grounded span in the source text, with its publication date.
The reasoning is blind to the market
The model that builds a causal chain is never shown a ticker, a fund, or your watchlist. It reasons in economics — feedstocks, capacity, demand, constraints — and vehicles are attached only at the end. A chain cannot be reverse-engineered from a symbol somebody wanted to sell you.
Counterevidence is looked for on purpose
Grounding reaches for reporting beyond your own feeds, and every chain gets an adversarial review whose only job is to surface what a tidy story leaves out. Where nothing can be found, the gap is recorded rather than quietly dropped.
Confidence is calibrated, never invented
Ratings are measured against realized outcomes. Where there isn't enough history to calibrate, the product says so rather than guessing.
Your watchlist is yours
Per-user isolation is enforced in the database itself, not just in application code. The expensive analysis is shared; your watchlist and prompts are not.
FAQ
Questions worth asking before you trust any of this.
What does Priors actually do?
Priors ingests the news sources you configure, extracts grounded events, then reasons outward from each event through the economy — one step at a time, with a stated mechanism, observable and falsifier at every hop. The model doing that reasoning is never shown a ticker or your watchlist. Each chain is then checked against dated evidence and counterevidence drawn from beyond your own feeds, and only the terminal economic effects are matched to investable vehicles. Those links are backtested against historical price data. What survives becomes a directional, confidence-rated signal; what doesn't is still shown to you, labelled as unvalidated research.
How is this different from an AI stock-picking tool?
An LLM is good at proposing obscure multi-hop causal links but cannot be trusted to score them. Statistics can validate links but cannot infer world-knowledge causality. Priors uses each to cover the other's weakness. The reasoning step is also deliberately catalog-blind: the model is never told which assets exist or which ones you watch, so it cannot quietly reason backwards from a ticker to a story that justifies it. Assets are attached afterwards, at the end of a chain that was built without them, and nothing becomes a signal until it passes statistical validation.
How are causal relationships validated?
Each proposed link is backtested against historical price bars using event studies over the relevant windows, with family-wise false-discovery-rate control so that testing many hypotheses does not manufacture false positives. Confidence is calibrated against realized outcomes, never fabricated. Only validated edges can appear inside a signal, and research that has not been validated is never given a confidence score.
Do I only see things that passed validation?
No, and that is deliberate. Validation is a hard gate on signals: an unvalidated edge cannot appear inside one. But a proposed causal chain that hasn't been tested yet, or that can't be tested because the price history is too thin, is still worth reading — so Priors shows it, always marked as a proposed hypothesis, a cross-source synthesis, or a coverage gap where no evidence could be found, and never dressed up with a confidence number. The label on every item tells you exactly which of those you are looking at.
Is Priors financial advice?
No. Priors is analytical software — not a broker, dealer, or investment adviser — and nothing it produces is a recommendation, solicitation, or an offer to buy or sell any security. The analysis is impersonal: it is computed once, globally, and is identical for every user. A watchlist only filters which parts of that shared graph you see, and the reasoning step is not shown your watchlist at all. Priors does not know your positions, financial situation, investment objectives, or risk tolerance, and never tailors output to them. Priors is not compensated by any issuer, broker, exchange, or fund for coverage. Output is directional and horizon-bounded — a direction, a time horizon, a calibrated confidence, and the causal chain that explains it. There are no price targets, no execution, and no guarantees. Every decision is yours alone.
What is a multi-hop causal chain?
A relationship that reaches an asset through intermediate steps rather than directly. A Strait of Hormuz disruption affects crude supply, which changes ethanol feedstock economics, which propagates into US corn prices. Priors traces those intermediate hops explicitly and shows you every one — including hops through assets you aren't watching, which is usually where the non-obvious part lives.
What data does Priors use?
News comes from the RSS and Atom feeds you configure — you choose your sources. Optional source packs let you install a reviewed bundle of public feeds in one step instead of adding them one at a time, and you can see every publisher in a pack before you install it. When Priors grounds a causal chain it also draws on public reporting from beyond your own feeds, so a chain is not confirmed only by the outlets you already read. Price history comes from market data providers including Tiingo, Twelve Data and Finnhub. Cheap local pre-filtering gates the expensive model step, so cost stays bounded no matter how much news you point at it.
Start with one watchlist and one feed.
Sign up, add the tickers you want to watch and the sources you actually read — or install a source pack in one click — and let the first run tell you what connects them.
Free during early access. No card required.