A headshot changes the true win probability the instant it lands. The market price needs longer — it has to be seen, believed, and then quoted on a one-cent grid by someone willing to hold the risk. This project measures that window on granular CS2 and Dota 2 game state, and asks what it costs to be the one providing liquidity across it.
State-based win-probability models, calibration and event-response latency, and a sandbox for inventory-aware quoting — built as an academic, noncommercial research prototype.
Illustrative reconstruction from a scripted round, not a measured result. The shape is the hypothesis the study is designed to test: belief steps on the event, price ramps toward it, and the spread flares while the maker prices what just happened.
An esports match emits structured events at high frequency: a kill with position and weapon, a round won on a specific condition, an objective taken, an economy that shifts by a known amount. Each one is a discrete, timestamped update to the state of the game — and therefore to the conditional probability that a given side wins.
A prediction market on that same match updates differently. It moves when a participant observes the event, forms a belief, and posts an order that someone else takes. That path has latency, it has disagreement, and on venues like Kalshi it terminates on a one-cent price grid, which means the finest expressible change in belief is a full percentage point of probability.
Between the event and the settled price there is a window. This project measures it — how wide, how it varies by event type and game phase, and whether a maker quoting through it is being paid for inventory risk or simply being picked off by someone with a faster read on the state.
Round score, player counts, economy and buy state, bomb/objective status, map control, and time remaining — fitted jointly, then ablated. The output is not just a model but a ranking of what moves belief, and by how much, conditional on game phase.
Aggregate Brier score hides the failure mode. Reliability is evaluated conditionally — post-plant, low-economy, late-round, high man-advantage — because a model that is well-calibrated on average and overconfident after a plant is a losing model precisely when size is on.
For each event class, estimate the response function of publicly available market prices: dead time before first reaction, time constant of the ramp, whether it overshoots, and how much of the total move has landed at 1s, 5s, and 30s.
Fills taken during the absorption window are adversely selected by construction. The question is whether an inventory-aware, event-aware quoting policy earns enough spread in the quiet intervals to fund the losses in the loud ones — and how much of the answer is destroyed by a one-cent tick.
GRID's official CS2 and Dota 2 feeds, normalized to a single timestamped event schema with explicit clock reconciliation between feed time and match time.
→ event tapeEvents folded into a running state vector: score, alive counts, equipment value, objective status, map control proxies, time remaining. One row per state transition.
→ state framesGradient-boosted and sequence baselines fit on state frames against realized round and match outcomes, with strict temporal splits and no post-outcome leakage.
→ p(win | state)Reliability diagrams and proper scoring by conditional slice; event-study estimation of the price response function against public market data.
→ diagnosticsBacktest of inventory-aware quoting and risk control against the reconstructed tape, with tick constraints, position limits, and adverse-selection accounting.
→ policy evaluationBoth titles reduce to the same shape: a small set of discrete, verifiable state variables plus a clock. That shared shape is what lets one modeling and evaluation framework cover both.
A win-probability model that says 80% should be right 80% of the time. Checking that on the full sample is table stakes. The diagnostic that matters here is conditional: the reliability curve restricted to post-plant states, to man-disadvantage retakes, to the final thirty seconds — the states where an automated quoter is holding the most inventory risk.
Reported alongside the curve: a Brier decomposition into reliability, resolution, and uncertainty, so that a model which is merely non-committal is never mistaken for one that is well-calibrated.
Illustrative shape, showing the failure mode under test: well-behaved mid-range, overconfident in the high-probability post-plant tail.
Share of the eventual price move completed since the event, by event class. Dead time, ramp, and overshoot are the estimated parameters.
Regulated event-contract venues such as Kalshi quote binary contracts between 1¢ and 99¢ in one-cent increments, where the price is read directly as a probability. That grid has a consequence most equity-market intuition doesn't prepare you for: the minimum expressible spread is one full point of probability, and it does not get finer when the market gets calmer.
So a maker faces two distinct regimes. When the theoretically optimal half-spread is wider than the tick, the quote is spread-bound and the usual inventory logic applies. When it is narrower — a quiet mid-round, a well-understood state — the tick binds, and the maker's only choices are to quote a spread wider than the risk warrants, join the queue at a price they don't love, or step away and forgo the flow entirely.
The sandbox below makes that regime switch explicit. Reservation price and half-spread follow the standard inventory-aware formulation, then get snapped to the venue's grid — which is where the theory stops being clean.
Reservation price r = p − qγσ², half-spread δ = γσ²/2 + (1/γ)·ln(1 + γ/κ), then snapped outward to the venue tick. κ is the fill-intensity decay; on probability-space prices it must be large, or the model quotes an absurd spread. Parameters are illustrative — estimating σ and κ from real event-conditioned data is part of the contribution.
Reproducible ingest, state encoding, model fitting, and evaluation, with pinned environments and deterministic seeds.
Round- and match-level replay of state, model belief, and public market price on one clock — the diagnostic surface for every claim in the work.
Backtest environment for inventory-aware quoting and adaptive risk control against the reconstructed tape.
This is a research prototype in support of doctoral work and a prospective scholarly paper. The commitments below are the terms the project operates under, stated up front rather than negotiated later.
The initial system is a research prototype: analysis notebooks, a visualization dashboard, and a backtesting sandbox. It is not a product and does not take positions in any live market.
GRID's raw data is not republished, re-shared, or exposed through any interface. It stays inside the research environment.
Anything published contains aggregated or derived results, subject to GRID's publication and attribution requirements.
Betting or live commercial use may require separate due diligence, written approval, and an appropriate GRID license. Open Access data will not be used for such activity without that authorization.
Whether this research could eventually support market-making tools for regulated prediction markets is an open question the project is evaluating — under a separate agreement, if and when it arises.
Price-response analysis uses publicly available market prices. No privileged venue data and no order-level information beyond what any participant can observe.