Lagisalpha

The lead-lag edge in prediction markets. Built on TxLINE.

A prediction market sets its price by trading, so it lags the sharp, vig-free line that already holds the true probability. TxLINE strips the bookmaker margin from a live World Cup odds feed to produce that true probability - and it moves the instant news lands. When a prediction market (Polymarket) falls below fair, the cheap side is underpriced. Lagisalpha detects that divergence, sizes the trade by Kelly, takes profit at fair, and proves the result on real on-chain fills.

It never takes the other side of a bet - it is a measurement-and-signal product, not a book.

Built for the TxLINE / TxODDS World Cup hackathon (Solana).

Technical documentation: see TECHNICAL.md for the full architecture, data flow, and API reference. The litepaper covers the thesis and the evidence.

Paper-trade it

The trader-facing product is a paper-trading terminal over the same signals. No install:

npx lagisalpha

Set a bankroll, pick a match, and watch each team's cheap side converge to TxLINE fair as a Kelly-sized paper trade with live PnL. Replay is open; live needs a key (load las_...). It is also on Telegram - @lagisalphabot - as alerts, or paper trades on a bankroll you set. Signal only, no real orders. Each call labels the team whose price is cheap; the trade is the convergence to fair, not a bet on who wins.

The edge in one paragraph

Work in probability space. TxLINE's de-vig 1X2 gives the fair probability a team wins; the market's moneyline gives its own probability of the same event. When fair sits above the market price by more than a threshold, the cheap side is underpriced and we mark an entry - which side, how far off fair, and how much size you could later exit into at fair. Buy the cheap side, take profit at fair when the market catches up, size each bet by Kelly on the gap. Holding to the final result is a losing trade on this data; the convergence is where the money is.

The proof

Measured on settled World Cup matches, on the real fills, over every call the detector fired - no exclusion filter, nothing curated:

• Reach - does the market price travel back to fair before the match ends? Currently ~79% of the time, recomputed live on /proof. Outcome-independent, so it is the firmer number.
• Return - Kelly-sized (capped at 30% per call), take-profit-at-fair, compounded across every call. Take-profit far exceeds holding to the final result; the convergence is where the money is. The compounded figure is concentrated at pilot size and published as-is on /proof, where it recomputes as each match settles.

Pilot sample (13 matches): the compounded return is carried by a few high-volume matches, so it is a pilot, not a promise. Reach is the firmer read; both tighten as matches accrue.

Signal policy. Every call counts: either side, any size, any minute, each side named by its team. Sizing is the only risk control: Kelly on the gap, f = gap/(1 - price), capped at 30% of the balance per call so no single bet can ruin the account (full Kelly, uncapped, once staked 81% on one call and gave back 76% of the bankroll). The maths is computed the same way on the box and the site.

What we found (pilot)

The brief floated a Sharp Movement Detector - flag significant TxLINE odds shifts and see if they call the result. We built it, found it is a coin flip, and did one better:

• Odds shifts alone: 58% (7/12) at calling the result - no better than chance. The line moving is not the edge.
• The lead-lag is. A goal is new information: TxLINE reprices it instantly, a prediction market only moves when someone trades, so for a window the cheap side sits below fair and converges ~79% of the time. It is our strongest, most proven signal, and the record rolls unfiltered: every call is published and scored, with the calls that hurt it left in.
• Goal-imminent alerts: a TxLINE high_danger_possession makes a goal by that team ~4x more likely within 2 minutes, and a divergence it flags converged 84% vs 75% without one.

A volume-to-divergence read (more traded money per point of divergence tends to mark the winner) rides along in the terminal as an experimental overlay, graded live and penalty-honest; we do not lean on it. In-sample on 12 matches; a promising pilot, not a settled result.

Architecture (short version)

TxLINE SSE (fair) + Polymarket fills (Polygon) → EC2 pipeline → Supabase blob → Next.js site. A Python pipeline on an EC2 box streams the de-vig fair line, decodes real Polymarket fills from Polygon, joins them, computes reach / return / Kelly every 30 min, and publishes desk-archives/pickoffs.json plus a replay index and one replay blob per match to Supabase storage. The Next.js app reads those blobs and renders the site - every headline number is dynamic, never hard-coded, and per-match replay data is served through CDN-cached routes (a finished match never changes). Full detail in TECHNICAL.md.

Verifiability

Both legs are public. The fair side is TxLINE's World Cup feed - odds and scores anchored on Solana, access minted by a real on-chain subscribe transaction (surfaced on /proof). The market side is real fills read straight from Polygon, decoded to a price and size per trade. Open any fill as a Polygon transaction, settle any outcome on TxLINE's on-chain scores, and recompute the edge yourself. Nothing here is asserted.

Endpoints

Public:

• GET /api/live-edge - live in-play divergences: { generatedAt, liveCount, theta, signals[] }.
• GET /api/replay-edge - same shape over the bundled replay matches.
• GET /api/replay-signals - per-match replay feed with entryFill/exitFill, goal-watch and winner-hint; powers the open npx lagisalpha replay and /launch.
• GET /api/live-stream - tick-by-tick TxLINE + Polymarket snapshot behind /live.
• GET /api/live-frames - real-time TxLINE frames (polled snapshot).
• GET /api/verify-csv - per-frame verification CSV for reconciliation against the provider.

Signal API (authed - Authorization: Bearer las_...; buy a key at /api):

• GET /api/v1/divergences - the canonical trader signal feed. ?status=live (gated to a live match, else no matches live), ?match=<fixtureId>&theta=5|10 (a settled match), or no params (match index). Each signal: side, team (the team whose price is cheap), entry, fair (take-profit target), gapPp, suggestedKellyF, sizeAtFair, ts.
• GET /api/v1/fair - current TxLINE de-vig fair per live fixture. We hold the TxLINE token and feed the fair, so a trader needs no TxLINE access of their own.
• GET /api/v1/track-record - pooled reach / Kelly ROI / CI plus per-match edge.

Retired (410 Gone): /api/v1/signals, /edges, /archive, /calibration, /control-room - the operator-era line-integrity surfaces.

Consumer / API pricing: USDC, chain-agnostic - $97.99 and $699.99 tiers.

TxLINE endpoints used

Access uses a server-held token (guest JWT + an on-chain Solana subscribe transaction → apiToken), sent as Authorization: Bearer <jwt> + X-Api-Token: <token>. The subscribe tx is the on-chain proof of access (/proof).

• GET /api/fixtures/snapshot - live fixtures, team names, kickoff times.
• GET /api/odds/stream - live de-margined (no-vig) odds (SSE) - the core signal input.
• GET /api/scores/stream - live scores + match events (goals / red cards, SSE).
• GET /api/odds/snapshot/{fixtureId} - current de-margined book, polled for the real-time frames panel.
• GET /api/scores/updates/{fixtureId} - full kickoff-to-FT score sequence, used to capture matches for the bundled replays.

Odds history is gated (/api/odds/updates is empty on the free tier), so the de-margined book is captured live off /api/odds/stream.

Configuration

| Env | Purpose | | --- | --- | | FEED_MODE | replay (bundled real captures, default), live (TxLINE streams), or synth (deterministic stand-in). | | TXLINE_API_BASE / TXLINE_JWT / TXLINE_API_TOKEN | Server-held TxLINE token (guest JWT + on-chain subscribe). | | TXLINE_SIGNUP_TX / TXLINE_CLUSTER | Solana subscribe tx + cluster - the on-chain proof of access shown on /proof. | | SUPABASE_SERVICE_ROLE_KEY | Server-only; the claim route appends issued las_ API keys (sha256-hashed) to the private desk-private/api-keys.json store, and the Signal API validates against it with the same credential. Nothing about the key store is world-readable. | | REPLAY_SPEED | Match-seconds per wall-second for replay mode (default 30). |

Develop

npm install
npm run dev        # http://localhost:3000
npm run typecheck
npm run build

License

AGPL-3.0.