Desk

Read it top down: the live flow, then what needs a decision, then the open risk book and the loss distribution behind it, then margin performance. Monitoring above, review below — the panels that say act now come before the ones that say how did we do. Click any match, selection or bar to open the inspector.

Bet flow movers only

Needs attention ranked by limit use all alerts ›

Risk & settlement book exact · no simulation

Event / market / selection	Odds	Prob.	Turnover	ABL	Exp. takeout	Liability	Exp. liab.	Exp. margin	Downside	Down. margin	Result it

Match	Result	Slips closed	Still riding	Stake (apportioned)	Realised so far	Still pending	Total contribution

Risk distribution & VaR simulated

before (open) after — profit after — loss VaR 95%

The simulation draws a scoreline, a corner score, and shots and shots on target per match — four coupled blocks, drawn from the shared factors first and then each block conditional on them, exactly as the closed-form integral does it — and settles every market off them. The latent state is sampled, not the markets. Reproducible: own generator, reseeded from the book seed before every run.

Book concentration

Margin on the book you are carrying every slip struck

theoretical expected

Margin on the book that has closed only the slips that have settled · all three on one base

Theoretical and expected are free; realized is not. The first two are running aggregates over arrival time — neither has a result in it — so they are exact and always drawn. Realised needs the clock walked, and walking the clock reassociates the engine’s calibration at the 1e-15 level: every value still prints the same, but it is not bit-identical, and a measurement that silently perturbs the book it is measuring is not something this console does. So the history line is traced on request.

Traced on a cleared board. A manual settlement carries no timestamp, so replaying the clock with one applied would show matches decided before they kicked off. This is the natural history of the book, not the history plus whatever is currently on screen. The dotted vertical is where the clock is standing.

Expected liability

Where the number comes from. Every bar is exact and additive — they sum to the book total, so this is the whole liability decomposed, not a sample of it.

By contribution exact · additiveopen book only

eventturnover ← downside · upside → net exp.P(loss)

Red is the probability-weighted value of every losing outcome, green the winning ones; the marker is the net. Sorted by downside, because turnover and danger are different lists. Open book only — money already settled is booked as realized P&L and has no expected liability left to decompose. Click a row to inspect the match.

Where it comes from area = exp. liability width = turnover · height = expected margin

Tail contribution simulated · additive · ±s.e.

Exposure lineage exact · additive

uncoupled — singles on this focal coupled — carried with another match node area ÷ ribbon width ∝ apportioned contribution hollow node = too small to size honestly hover for the numbers

Carried by	Slips	Apportioned	Share	Co-carried

carried by multis carried by singles build-up across the card

Alerts & limits

Exposures ranked against their ceilings, not by size — and against the right ceiling. A big number inside its limit is not a problem; a mid-size one over its limit is; and a huge number at 1% is a cheque you must be able to write, not a position you must trade out of.

Breach worklist act here

Scope	Match	Market / selection	Odds · prob	Exposure	Limit	Utilisation	State	Act on it

Utilisation is exposure ÷ ceiling at seven scopes in two families. Expected — selection, market, event downside over the joint grids, and book VaR 95% — weight liability by the probability it lands, and they rank this list. Payout — selection and market — are ceilings on what you can be asked to hand over, independent of probability. Coupling concentration is a share rather than an amount, because what it measures is concentration, not size. A payout breach below the plausibility floor is tagged longshot and sorted to the bottom: still counted, never hidden, because the obligation is real even when the price is not. Every row carries both numbers and the probability, so no row needs interpreting. Ceilings are set on Setup and auto-calibrated to the current book when you regenerate. Trace any row to see where its exposure came from. This panel does not settle. A book is settled by resulting the match — enter the score and every market on the card follows — and market-level settlement is an override that lives in the Event Inspector. Inspect opens it.

Liability by risk level book shape

profit if it wins loss if it wins open book (before scenario) equal expected liability dot size = turnover · hover to identify

Book shape, not a worklist: is the red sitting on likely outcomes or on longshots. The liability axis is signed-log, linear through zero up to a knee at the typical position and logarithmic above it. A book's liabilities span three orders of magnitude — a median position around a hundred euros and a longshot correct score in the tens of thousands — and on a linear axis nine tenths of the book collapses onto the centre line. Vertical distance is therefore not proportional to euros, which is why the axis carries real ticks rather than only its endpoints. The dotted hyperbolas are iso-expected-liability curves — points on the same curve cost the same in expectation, which is otherwise invisible because expected liability is the area under a point, not its height.

Top exposures worst open liabilities

Match	Sel.	Market	Odds · prob	Liability · expected	Limit use	Inspect · trace it

Scenarios & what-if

Move the clock, replay the card, or force outcomes across the whole book. Everything is reversible — the header shows when a scenario is applied.

Card clock 72h pre-match window

Rewind · replay

Result the card primitives in, markets out

Match	Result	State

This is the whole input surface. Four numbers settle a football match, two plus one conditional answer settle an ice hockey match, and every market on the card is derived from them. That is how a book is actually settled: a trader enters a result and the platform grades the markets, rather than grading market by market. The percentage beside each result is the model's probability for it, so a what-if shows whether you are pricing a coin flip or a one-in-fifty world.

Void is a state of the match, not a separate action. Voiding keeps the result you entered rather than discarding it, so un-voiding brings it back and you can see what you are setting aside.

Reverse scenario start from the outcome, find the card

The search space is not the grid. A card of sixteen matches has more scorelines than can be enumerated, but the book's P&L does not depend on the scoreline — it depends on which selections won. Two scorelines that grade every market the same way are the same world to this book, so the space to search is the set of achievable settlement signatures per block, which is small. That collapse is exact, not an approximation.

Two different questions. The worst card is a coordinate descent: sweep one block at a time, take whichever signature hurts most, repeat until nothing moves. Sweeping matters because multis couple matches together and one pass is not enough. The second search is the one a trading director actually asks — not what is the worst case but what is the most ordinary-looking Saturday that costs me this much — so it starts from the most likely card and repeatedly makes the change with the best ratio of loss bought to probability spent.

Both are local searches and are labelled as such. Coordinate descent finds a local worst, not a proven global one. On a card small enough to enumerate exhaustively the two agree, which is the gate this version ships with; on a full card it is a strong lower bound on the damage, not a certificate.

Scenario library stored as results, not as settlements

Scenario	Matches	P(settles this way)	Book P&L	Percentile	Where it sits

A scenario describes the open book. A match that has kicked off has a result, and no bulk driver, generator or search may rewrite it: they are all statements about what is still to play. Typing into a played match's row by hand still works, because what if that had gone the other way is a real question — but it has to be a deliberate act, not a side effect of pressing a button labelled most likely. Deliberate overrides are stored with the scenario; incidental ones are not. The probability column counts the open matches only.

The standing six are generated, not stored. A saved scenario is keyed to a card, so a library shipped with fixed results would go stale the moment you regenerate. These six are rules — every favourite wins, the coupled pair turns over, the two biggest matches are called off, and three searches — so pressing the button after a regenerate asks the same six questions of the new book. Anything you save yourself is left alone when they reload. The two thresholds are expressed as a share of handle rather than a euro figure, because €25,000 means one thing to a state operator and something else entirely to a start-up.

A scenario is an outcome vector. One cell per latent block per match — exactly the shape the simulator draws and the pricer prices. A scenario you typed and a scenario the model drew are therefore the same object, which is why both can be priced by the same function and put on the same axis. The percentile is where a named scenario sits among worlds drawn from the calibrated grids: a −€40k result at the 3rd percentile is a different conversation from the same number at the 30th.

Read the probability column carefully. It is the probability that the card settles this way, not that it produces these exact scorelines. Two scorelines that grade every market identically are the same world to this book, so their probabilities are summed — which is the figure a trader wants and is orders of magnitude larger than the per-scoreline one. It is still small: a fully specified card of sixteen matches is a rare world however it is counted, and that is the reason the percentile column exists. No individual world is likely, so the question worth asking is not how probable this one is but how bad it is relative to the ones the model would draw.

Why results and not settlements. Market outcomes are derived and lossy — forty-five football scorelines produce the same five settlements — so a library of them could not be replayed. A library of results can, and it survives a regenerated book. Where the card has changed underneath a scenario, the row says so rather than applying it silently. Matches a scenario does not cover take their actual result, and the coverage count is beside the name.

No VaR or ES column, deliberately. Those measure the tail of an open book. A fully resulted scenario has no tail left: it is one world, priced. Putting a tail figure beside it would be inventing uncertainty that the scenario has already resolved.

Force outcomes across the book reversible

Forcing an outcome conditions the latent grid onto the surviving state space rather than collapsing it to a point, so partially settled matches keep pricing correctly on the markets still open.

Why the two rows are different. The selects above force one market across the card. That is how a trader thinks and it is the right control for a what-if, but a market is an instrument: settling instruments one at a time can leave the underlying state undetermined, and the match stays open even though every chip on its row has been pressed. The row below settles from a state instead — one achievable final score and tie-break per block, every market read off it, including correct score and margin, which have too many selections to sit on a row. That is the scenario definition a market risk system uses: shock the factors, revalue everything. It is complete by construction, and the residual is asserted on screen rather than assumed.

Slip ledger ordered · replayable

#	Arrived	State now	Slip	Detail

Every slip in arrival order, with where it currently stands. The whole book is rebuilt from this ledger on every clock move, which is what makes a scenario reversible. Leg resolution is derived from the clock rather than stored, so it is recomputed on replay instead of appearing as its own ledger entry — a production deployment would persist those events too, and that is what turns this into a full audit trail.

Model & calibration

The engine room. One latent state per match — four coupled grids for football, a regulation scoreline plus a tie-break channel for ice hockey — calibrated to the probabilities the platform already holds, or recovered from quoted odds where it does not. betRISE never sees a price book.

Latent grids marginals in, joint distribution out Match:

Four latent blocks per match. The goals grid (home×away goals) spans 1X2, double chance, O/U and BTTS. The corners grid spans the corner totals, most corners and the corner handicap. The shots and shots on target grids span their totals and most-shots markets. Correlation inside each block is exact, and the blocks are coupled to each other rather than combined as independent — they share two factors and are tied on the latent scale by a Gaussian copula. The cross-block error above is what treating them as independent would cost instead.

Margin discovery recovered from the flow, not configured

Market group	Measurable	Coverage	Recovered	Configured	Error	Drift

This is the production shape. An operator configures margin in their own platform and betRISE never sees the configuration — it sees the flow. The estimator is the over-round itself: for a market whose selections partition the outcome space, the booked probabilities sum to it. The Configured column exists only here, in the demo, so the recovery can be checked against the truth. On a real feed there is nothing to check it against, which is exactly why the coverage and drift columns matter.

Two kinds of market cannot be read this way, and both were found by measuring. Fold sources — double chance covers the outcome space twice, so summing 1X, 12 and X2 double-counts; included, the three-way group recovered 1.229 against a configured 1.060. The engine already knew: ingest() skips fold sources when calibrating, for the same reason. Whole lines carry push mass, so their selections sum to less than one — Asian 2.0 covers 0.714 of the space. Both are excluded and counted rather than quietly averaged in.

Coverage is the honest constraint. betRISE learns a market's prices only from bets struck on it, so a market is measurable only once the flow has revealed every selection. Two-way markets fill quickly; a correct score with twenty-two selections almost never does. Drift is the widest gap in snapshots between the observations that make up one estimate: prices move across the window, so a market assembled from distant observations carries that movement in its over-round.

What this engine does not price, and why the boundary

Every market above is a function of a final state, which is what makes it a projection of a grid and what makes correlated exposure computable rather than asserted. A market that settles on the order or the timing of events is a different object. Naming them is not a disclaimer: it is the line that separates what has been built from what would need a within-game process model, and it is the same line tennis sits on.

Not priced from the grid	Why

Contribution to Expected Shortfall simulated three limits, on screen

The book's ES 95% is one number and the console splits it by match. The split is the Euler allocation, which for expected shortfall is a conditional expectation — a match's average P&L across the book's worst 5% of draws — and it is the split rather than one of the alternatives because it is the only one that adds up: expected shortfall is homogeneous of degree one, so by Euler's theorem the contributions sum to the total exactly. That is asserted on every build, not assumed.

Limit	What it means here
Zero is not safe	A match that never lands in the book's worst 5% contributes exactly zero, however much it could lose on its own. This is the known defect of every threshold allocation — Bodoff's objection is that loss scenarios below the threshold use capital and are allocated none of it — and it is mitigated rather than solved: the standalone column beside every contribution is what that match can lose on its own bad day, and the diversification index is the ratio of the two.
It carries a standard error	A contribution is a mean over the tail sample alone — 750 draws of 15,000 at the default — not over every draw. Expected shortfall already needs a larger sample than VaR for the same accuracy; a contribution needs more again. Every figure is printed with its ± and none is quoted to more precision than that supports. Raising Monte-Carlo samples above tightens them.
Two apportionment bases, on purpose	A multi's P&L is split across its legs by net-win weights. Realised and pending P&L use every leg, because a settled leg did contribute its share of what was made. Tail contribution uses the open legs only, because scaling exposure to a match that has already resolved changes the draw not at all — its contribution is identically zero. Different questions, different bases; both are correct for their own question and neither is correct for the other's.

Realised money never enters a contribution. The simulation walks only open slips, and the book's banked P&L is added identically to every draw — a constant cannot change which draws are in the tail and is never split across matches. It appears on exactly one line, the total, where the exposure page reconciles still at risk + banked = ES 95%. Ceilings are set on the still-at-risk figure, so banking a profit on one match cannot loosen the tail ceiling on another.

The counterfactual in the event inspector — what the tail would be if this match had never been written — is a different measure and is offered as a button rather than a column for two reasons: it costs a full simulation per match, and it does not add up, because removing a match also removes the diversification it was giving everything else.

Singles and multis simulated one book, decomposed

Operators run separate ceilings for singles and multis at every level — selection, market, event, competition, sport, book. betRISE deliberately does not, with one exception, and the reasoning is worth having on screen because it is the opposite of the industry's.

Where	What betRISE does, and why
Risk measures expected liability, downside, tail	Decomposed, never split. A multi's risk sits on the same matches as the singles — the console apportions it there and the sums reconcile exactly — so "the multi book" is not a book, it is a way of arriving at exposure that is already counted where it lands. Splitting the ceilings would also import the problem the industry grid has: exposure limits multiply at higher levels of aggregation, which is what two bet types times six levels is, while a unified measure gives one limit that works at every level.
Payout ceilings solvency, not risk	Split, at selection level. The worst single on this card pays €1,513; the worst slip of any kind pays €62,917. One ceiling across a 42× gap is either too loose to bind on a single or too tight to allow normal multi business. Same argument the Basel large-exposures framework makes for keeping a crude single-counterparty cap beside risk-based capital: a portfolio measure is an average over everything and will not see the one position that hurts on its own.
Bet-type concentration	One ceiling on a ratio, replacing what a second grid of amounts would have done. Not how much turnover is in multis — that tells you almost nothing — but how much of the tail sits in one half of the book, so drift in that ratio raises its hand.

What the split actually shows. Singles hold a diversification index near 0.2–0.4 all card — a hundred singles on different matches largely cancel — while multis run 0.95 to 0.99 under every mix tested: the multi book is the tail. And the ratio moves. As multis accumulate won legs their surviving stake gets more leveraged and more correlated, so singles' share of the tail falls steadily through the evening while their turnover share barely moves. A pair of ceilings set before kick-off is watching a number that changes by a factor of five during the card.

And it is concentration, not volume. Reaching a 65/35 turnover split with more small singles puts singles at about a tenth of the tail; reaching the same split by making the existing singles two and a half times bigger puts them at a third. Same ratio, four times the risk. That is why the instrument that controls a singles book is a payout cap and not a risk limit — it caps the thing that actually drives the danger.

Method changes propagate through the whole book

betRISE de-margins with Profile B only Unused when the betslip carries the platform's fair probability. Where it does not: net-win stable is parameter-free, round-trips to machine precision, and inverts a 50-way market as cleanly as a 2-way.

Monte-Carlo samples Used for VaR, ES and the per-match ES contributions. The expected liability is closed-form and cross-checked against the draw. Raising this tightens the standard error on every contribution, which is a mean over the worst 5% of these draws rather than over all of them.

Calibration slack 0 when the probabilities are supplied or the source's pricing method is known. Raise it when de-margining an unknown third-party feed.

0.0

Source prices with Stand-in for the operator feed betRISE is reading. Changing it regenerates the book.

Setup

Operator configuration. Set once, not per session — unlike the book generator, which is demo data and lives in the drawer.

Exposure limits drives every alert

Expected liability · selection ranks the worklist Liability weighted by the probability it lands. This is the economic control: what the position costs, not what it could cost.

Expected downside · market Sum of liability × probability over the outcomes that lose. The same measure as event downside, one level down.

Max payout · selection a ceiling, not a risk measure Most you can be asked to hand over on one selection. Independent of probability on purpose: if it lands you owe it whatever the price said.

Max payout · selection · singles only the one ceiling that splits by bet type The same measure as the ceiling above, over the singles book alone. This is the only place betRISE splits a ceiling by bet type, and it is deliberate: on the reference card the worst single pays €1,513 while the worst slip of any kind pays €62,917, a factor of 42, so one ceiling across both is either too loose to ever bind on a single or too tight to allow normal multi business. It is the same argument the Basel large-exposures framework makes for keeping a crude single-counterparty cap beside risk-based capital — a portfolio measure is an average over everything and will not see the one position that hurts on its own. The risk ceilings are not split, because a multi's risk sits on the same matches as the singles and betRISE decomposes it there instead.

Max payout · market Worst single outcome across a market's selections.

Plausibility floor payout scopes only A payout breach under this probability is tagged longshot and sorted to the bottom of the worklist — still counted, never hidden. Without it the list ranks by lottery-ticket size: measured on three books, the median probability of a firing payout alert was under 3%, and four of the top five rows were correct score. With a hard cut-off instead of a demotion the same books go from four alerts to one, which is the opposite failure. 2.5% is a starting point, not a standard.

Event downside Probability-weighted downside over the joint grids for one match.

Event contribution to ES 95% simulated Ceiling on how much ONE match contributes to the book's expected shortfall — its average P&L across the worst 5% of book outcomes, which is the Euler allocation of ES and therefore sums across matches to the tail itself. Every other event-level ceiling here is a ceiling on an average; this is the only one on the tail, and the two rank differently enough that the worst expected liability on a card is routinely not the worst tail contributor. Set on the open-risk basis: realised money is a constant in every draw and is excluded from both sides, so banking a profit elsewhere cannot loosen this ceiling. The figure carries a standard error because it is a mean over the tail sample, not over every draw.

Book VaR 95% Ceiling on the 95th-percentile loss across the whole open book.

Bet-type concentration simulated a share, not an amount Largest share of the book's tail sitting in one bet type. Not turnover — turnover tells you almost nothing here. On the reference card singles are roughly half the money and, depending on how concentrated their stakes are, between 3% and 38% of the worst 5% of outcomes; multis run a diversification index of 0.95 to 0.99 under every mix tested, which is to say the multi book is the tail. This ceiling exists so that drift in that ratio raises its hand, and it replaces what a second grid of single/multi risk ceilings would have done — one number on a ratio instead of twelve on amounts.

Coupling concentration a share, not an amount Largest share of one match's expected liability that is carried by slips also riding on a single other match. The only ceiling here that is a percentage, because the risk is concentration rather than size. Judged against an effective figure: the base share scaled up by as much as 50% when the coupled exposure sits in one market on the other match rather than spread across several, since one outcome then resolves all of it together. The scaling is a judgement, and both figures are always shown.

Warning threshold Utilisation at which an exposure turns amber.

75%

Hierarchical, the way a trading desk sets them: selection inside market inside event inside book. Utilisation is tracked at every level at once, so a book that is comfortable overall still surfaces the one selection that is not. Coupling concentration sits outside that nesting on purpose: it is not a bigger container, it is a different question, and it is the only one here that a conventional screen cannot compute at all.

And one that is neither. Event contribution to ES 95% is not a bigger or smaller container than the ceilings around it — it is the same container, the match, measured in the tail instead of in the mean. It exists because the two orderings disagree, and they disagree at the top. On the reference card the worst match by event downside — the ceiling directly above this one — is not the worst match in the tail, and the two rankings correlate at ρ = 0.59; against the expected liability decomposition on the exposure page the correlation falls to ρ = 0.22. The match that led the tail carried 32% of the whole open tail on its own while sitting third by expected liability. A desk that only watches the mean does not see that match until the night it happens. Both numbers are properties of one synthetic card, not constants — what is structural is that a mean and a tail rank differently, not by how much.

Two families, deliberately. The expected ceilings are risk measures — liability weighted by whether it happens — and they rank the worklist. The payout ceilings are solvency controls and answer a different question: can this book pay. A desk runs both, because a notional cap does not care how unlikely the outcome was. Event downside, book VaR and coupling concentration were already probability-weighted; selection and market were the two that were not, which is why the worklist used to open on correct-score longshots.

Feed & ingestion

Betslip profile A: the record carries the platform's fair probability, so betRISE uses it directly and no margin is stripped. B: odds only, betRISE recovers the probability itself. A is the common case on any platform that supports cash-out. The grids are calibrated either way, because a joint distribution cannot be read off marginals.

Alert routing Where breaches are pushed.

Auto-suspend on breach Pull the selection when it crosses its ceiling.

Base currency

Only the betslip profile and the limits are wired in this proof of concept. The rest is shown to place the settings that a live deployment needs, not to imply they are implemented.

betRISE risk console v72 · Logomath Analytics · synthetic data, no operator feed · figures illustrative