OptionsMath AI is a thesis-to-options solver: a language model turns your market thesis into explicit, editable scenario assumptions, then a deterministic options engine prices every candidate structure across those scenarios and ranks them by modeled expected value under the assumptions you can see and change.
Why the language model never touches a price
Language models are good at reading a market view and bad at arithmetic they have to invent.
Ask a general chat model for an options trade and it will give you strikes, premiums, payoffs, and a confident ranking. Those numbers are generated the same way its prose is — as plausible-looking text, produced in one pass, different on the next run. The failure mode is invisible, because wrong option math looks exactly like right option math.
OptionsMath AI is built so that failure cannot occur: the language model is confined to the job language models are actually good at, and it is structurally unable to do the rest. It emits a constrained schema and stops. Every price, payoff, penalty, and rank after that comes from compiled numerical code that produces the same answer every time it is given the same inputs.
-
Language model
01. Your thesis
Plain English, with a ticker. One sentence is enough.
-
Language model
02. Explicit scenarios
Branches with horizon, move, volatility shift, and probability — editable before anything is priced.
-
Deterministic
03. Candidate search
Structures enumerated across 29 strategy families against the live option chain.
-
Deterministic
04. Pricing tensor
Every leg priced in every scenario by stochastic models calibrated to that chain.
-
Deterministic
05. Ranked deck
Candidates scored, penalised, ranked, and shown with the assumptions that produced them.
Exactly what each side does
Writes the forecast
- Reads your thesis plus the live spot and as-of date for the ticker.
- Returns a constrained schema: an objective, and one branch per outcome with horizon, move, volatility shift, skew shift, and probability.
- Branch probabilities are normalized to sum to 1 and clamped to a valid range before the engine sees them.
- If the model fails or returns nothing usable, a deterministic parser produces the scenario tree instead.
- It never prices an option, picks a strike, or ranks a structure.
Does the options math
- Enumerates candidate structures across 29 strategy families, from single legs to four-leg condors, calendars, and diagonals.
- Calibrates an SSVI volatility surface and the structural models to the live option chain, and records each model's fit error.
- Prices every leg of every candidate in every scenario, using only the models whose calibration error is within 4x the best fit.
- Computes modeled P&L per scenario, then probability-weighted expected value, probability of profit, expected shortfall, and capital at risk.
- Subtracts an explicit assignment-risk penalty and a model-disagreement penalty, then ranks what survives.
score = probability-weighted modeled P&L − assignment-risk penalty − model-disagreement penalty
Both penalties are reported per candidate, so any ranking can be checked against its own inputs. A structure with short legs that could be assigned early is charged for it; a structure whose price the surviving models disagree about is charged for that too, rather than having the disagreement averaged away.
One thesis, 1,264 ways to express it
This is an actual analysis, not an illustration. The thesis was: “META upside into a model release and a pre-earnings run-up, with a branch for a delayed catalyst.”
The language model turned it into 6 explicit branches for META as of 2026-05-20:
| Branch | Probability | Horizon | Move |
|---|---|---|---|
| Core model release | 32% | 2mo | +20% |
| Early product run | 18% | 1mo | +10% |
| Delayed release, post-earnings | 15% | 3mo | +20% |
| Delay fade | 13% | 3mo | +5% |
| SOTA surprise | 12% | 2mo | +30% |
| Capex or earnings disappointment | 10% | 3mo | −10% |
The engine then fitted a volatility surface to 81 live quotes on that chain, enumerated and priced 1,264 candidate structures across all 6 branches, and ranked the ones that satisfied the constraints. The interesting number is not any single modeled payoff — it is that a one-sentence view has 1,264 plausible expressions, and comparing them by hand is not something a person does.
Change one probability in that table and the ranking below it changes. That is the correct behaviour: the deck is an answer to these assumptions, not a prediction that any of them will happen.
Precise math on a wrong forecast is still wrong
The engine calculates precisely under the assumptions it is given. If the scenario tree is wrong, the result is a precisely calculated analysis of a world that does not happen. No amount of numerical care fixes that, and nothing on this page claims otherwise.
This is why the branches are shown and editable rather than buried behind the answer. Reviewing them — and disagreeing with them — is the part of the work that stays yours. The tool's job is to make your assumptions explicit and then exhaustively price their consequences, not to have the view for you.
Read the pricing, calibration, and scoring detail in the methodology, the scope and limitations in the FAQ, or how this differs from chatbots and traditional optimizers in the comparison.
Rankings are model-derived under the assumptions shown. They are hypothetical research outputs, not investment advice, recommendations, or promises of return.
The plain answers, in one table
| Product category | Thesis-to-options solver (AI options strategy research software) |
|---|---|
| Intended user | Options traders and researchers who have a market view to examine |
| Input | A natural-language market thesis naming a ticker |
| Language-model role | Interprets the thesis into a constrained scenario schema; never prices, picks strikes, or ranks |
| Quantitative role | Chain search, calibration, scenario pricing, expected-value math, penalties, and ranking |
| Output | A ranked deck of candidate option structures with per-scenario modeled P&L and full assumptions |
| Ranking basis | score = probability-weighted modeled P&L − assignment-risk penalty − model-disagreement penalty |
| Strategy coverage | 29 families, 1–4 legs: singles, verticals, butterflies, backspreads, ratio spreads, straddles, strangles, condors, ladders, calendars, diagonals |
| Asset coverage | US-listed underlyings with listed option chains; one underlying per analysis |
| Market data | Delayed public quote snapshots; every analysis records its as-of timestamp and quote provenance |
| Pricing models | SSVI surface plus Bates, Heston, Merton jump-diffusion, and Variance Gamma, calibrated per run with recorded fit error |
| Brokerage | Not a broker. No order routing, no execution, no custody of funds |
| Advice | Not investment advice. Outputs are model-derived, hypothetical research artifacts |
| Last verified | 2026-08-05 |
Bring a thesis
Describe a move in one sentence and inspect the scenario tree it produces before anything is priced.