ScalarDelta is a research engine for professionals who trade the tail of small-cap equities. We built our own corpus of company-level mechanism evidence, trained our own models on it, and score every name on the shape of its forward distribution — not on a price target.
Three tiers, monthly, cancel any time. The call is with a technical trader — we have nobody in sales.
Everyone has prices. Everyone has fundamentals. What almost nobody has is a structured, point-in-time record of why a small company's prospects changed — machine-readable, comparable across thousands of names and years, and built specifically to be tested against forward returns. That is what we spent the last years assembling.
Filings, 8-Ks and company news land in our own store every session. Each document is turned into typed mechanism evidence: what changed, in which direction, how material it is, and over what horizon. Not a headline count, not a sentiment score — a structured claim about the business that can be tested against what the stock did next.
Those evidence vectors cluster into mechanism centroids — the recurring economic storylines behind re-ratings: capacity coming online, a pilot converting to a contract, a decline turning terminal. The clustering, the centroid geometry and the scorers on top are ours, fitted to our corpus. The large language model does one narrow job: reading a filing into structure. The asset is the dataset and the fit.
Every name carries a daily state vector — convexity (BCS, SCS,
p_down), compression, position in its 52-week structure, cohort-relative
momentum, plus fundamental sleeves for pre-inflection and compounding profiles. A great story
in the wrong state is not a trade, and the engine is built to say so.
If you have built signals yourself, you already know where they break: pooled embeddings, survivorship, a mean that hides the distribution, and a backtest tuned until it is true. Here is how each of those is handled — and where we are still wrong.
Mean-pool a company's embeddings into one vector and every cluster collapses to the same point — the signal disappears while the dashboard still looks healthy. Each piece of evidence is matched against the centroids on its own document date. The entry fires on the narrative event, not on a company-level average.
Hit rate is the least interesting number we compute. The objective is the shape of the forward distribution: right-tail rate, left-tail exposure, average maximum adverse excursion, net over max drawdown. A 46% hit rate with a fat right tail beats a 62% hit rate that bleeds — which is exactly the trade small-cap specialists are already making by instinct.
The validation universe is frozen and includes the dead: 695 securities, 95 of them delisted, deliberately retained. Backtesting on names that are still listed today is backtesting with tomorrow's information. Every persisted result records which universe it ran on and how many names resolved.
Every playbook carries an effective sample size, a lower confidence bound and a deflated significance measure after multiple testing. Thin results sit in probation instead of in a brochure. On a call we will show you the ones we paused, and why.
A worked example from our own research: on volume-backed activations in pressed small caps, 24% of names never reclaim the event-day close and average −13.5% over the next 60 trading days. So activation is a watchlist event, and entry is gated on reclaim. You get the entry status, not just the score.
A separately validated reversal engine: a −3% intraday flush, in a name that has not outrun the index by more than 2% over the prior five sessions, with the macro gate evaluated at the fire minute, flat by the close. Cohort-specific and walk-forward tested. We will also tell you what did not survive — the news veto did not, and a −2% trigger is frequency, not edge.
Start free and read the output for a few weeks. Plug into what we already compute when you want it in your own stack. Put your own names through the engine when our universe is not your universe.
See what the engine surfaces before you decide whether to plug into it.
No drip sequence, no upsell mails. One email, unsubscribe by replying STOP.
Use what is already there: our universe, scored daily, over a REST API.
Monthly. Cancel any time. Your overlay, your execution — we do not touch either.
Your names, run through the same engine as ours, under a fair-usage policy.
Monthly. Fair-usage terms below — they are a compute constraint, stated plainly.
Priced in USD, invoiced monthly, cancel any time. No setup fee, no onboarding fee, no sales commission — there is nobody here to pay one to.
Nominated coverage is priced against compute, not against how much you might make from it. So the limits are stated up front rather than buried in a contract.
We have no people in sales. Book a call and you get a technical trader: someone who has run the backtests, knows which cohorts the engine is wrong about, and will tell you unprompted when a sample is too thin to believe. Bring your hardest question about the methodology — that is the entire meeting.
The free digest starts immediately. For the two paid tiers we answer personally, usually within a session or two, because a trader reads every one of these.
Everything that follows comes from a trader, not a sales rep. If you need something sooner, mail desk@scalardelta.com directly.
No. The digest and the API name candidates, the state that surfaced them and an entry status. We never give a target price or a position size. You are the professional; the engine is an instrument, not a manager.
No. A commodity language model does one narrow job — turning a filing or a news item into typed, structured evidence. Everything that produces alpha is ours: the corpus, the clustering into mechanism centroids, the convexity scorers, the classifiers and the fusion, all fitted on data we assembled. The dataset and the fit are the product; extraction is plumbing.
On a call, in detail: per-playbook walk-forward and out-of-sample tables, effective sample sizes, drawdowns, and the playbooks we paused. We do not print a headline return number on a marketing page, because a headline return from an in-sample backtest is worth nothing and you know it.
US listings are deepest. The intraday layer additionally runs live across European and Hong Kong sessions. Nominated tickers can be any listed equity we can source clean data and filings for — if we cannot, we will tell you before you pay for it.
They are ingested and scored for you, with their full available history backfilled on entry. They stay out of the public digest. Drop a name and it drops out of active reading.
We store your contact details to answer you and, for the digest, to mail you. Your ticker list is used to run coverage for you and nothing else. No list sharing, no resale, no tracking pixels.