Model the Game.
Engineer the Market.
We derive our own event and market datasets, weighing recent form against longer-run baselines to test for recency bias. Fair-value models meet live quotes; sharp-line movement and order-book pressure serve as institutional-flow proxies. Fees, liquidity, and risk govern which apparent gaps reach execution review.
Get the data in order first.
Box scores, fight results, lap data, odds snapshots, and live feeds arrive in different shapes. Each sport's pipeline normalizes its sources and attaches capture times to what it stores, so later analysis can account for what was known at the time.
Turn outcomes into fair values.
Sport-specific models and Monte Carlo simulation produce outcome distributions. Those probabilities become fair-value estimates, checked through calibration and walk-forward tests to reveal where a price holds up and where it needs work.
Reprice against the live market.
Streaming order-book changes and sportsbook lines update the market view as prices move. Pricing logic compares those changes with fair-value estimates, while freshness and risk checks flag stale quotes and adverse moves for review.
Turn a pricing gap into a controlled decision.
Alerting and execution research have distinct roles. Live monitors surface fair-value gaps and adverse moves. Separate order-research tools attach inputs and risk limits to quote candidates, then hold them for explicit approval. That separation keeps a model signal from becoming an unreviewed order.
Built below the surface.
The work connects historical data and simulation to fair-value pricing, live market state, and guarded automation.
Model the full range of outcomes.
Simulation builds a distribution of possible results for an event rather than a single pick. Those distributions become fair prices, and the prices are scored against later results in time order, so a model is judged only on what it could have known.
Engineered for the repricing loop.
Automated trading systems need more than a model score. We develop the components around fair-value pricing, streaming market state, quote-risk checks, and controlled order workflows. On streaming feeds, book updates arrive without waiting for the next full poll, reducing avoidable latency and repeated work in live pricing research.
Keep market state current
Order-book snapshots seed in-memory state; level-by-level deltas update what changed. Sequence checks detect gaps before an old book is treated as live.
Calculate fair value
Model probabilities and no-vig reference prices provide separate inputs for evaluating a market. Current quotes are checked against edge, fees, and the latest available context.
Prioritize repricing
Queue position, fill likelihood, adverse flow, and resting-order risk help surface the quote changes that matter first, instead of treating every tick as equally urgent.
Keep execution controlled
Live monitoring raises alerts; it does not place orders. Separate trading-system research stages quote candidates with risk limits and explicit approval before any action.
Modeling markets in motion.
An illustrative look at fair value, market response, and the gap between them through scoring changes. Liquidity and execution conditions are outside the view.
The result is only part of the record.
Keep the data snapshot and source timestamps with the run, so it can be reviewed as of that moment.
Write the assumptions and decision criteria down next to the output they produced.
Go back and compare the decision with what the event and the market actually did.
A model should be inspectable from the data it saw to the decision it produced.
The discipline guiding our data and model research.From the archive to the live board.
Start a conversation.
For partnership requests, research collaborations, and general inquiries, send us a note.
hello@dugoutx.com