Methodology

How the Model Works

ClearContent Intelligence uses eight quantitative models grounded in peer-reviewed academic literature to identify potential mispricings on Polymarket before informed traders correct the price.

EARLY VALIDATION PHASE
This model stack is in active validation. We have 13 resolved predictions. Statistical significance at our target accuracy level requires approximately 80–109 resolved predictions. The World Cup 2026 (June–July) will be the primary calibration event, adding 22 predictions. All performance claims should be read as early-phase observations, not proven durable edge. Every prediction is timestamped publicly before the event.
01 — THE CORE INSIGHT

Markets Open at 50 Cents

When Polymarket creates a new sports market, it opens at approximately 50¢ — implying 50/50 odds — regardless of the actual probability. Argentina versus Haiti opens at the same price as France versus England, even though our model gives Argentina a 97% chance of winning.

This mispricing is not random. It is structural. Polymarket's automated market creation does not incorporate team ratings, recent form, surface advantage, or any domain-specific signal. The price reflects nothing but the act of market creation.

Our WebSocket listener fires within seconds of a new market appearing. The scanner runs every five minutes as a backup. We get in before the market price adjusts to reflect reality.

THE STRUCTURAL EDGE

From 13 resolved predictions, every time we identified a market opening far from our model probability, the model was correct 12 times. The one miss (Lakers over Rockets) was a close game at 60.5% model probability — not a high-confidence call.

02 — THE MODEL STACK

Eight Integrated Models

M1 + M2
Sports Elo + Margin Adjustment
Bradley-Terry Elo model across 64 football nations, 42 UFC fighters, 27 tennis players, 20 cricket nations. M2 adjusts ratings using margin of victory rather than win/loss alone — a 5-0 win updates Elo more than 1-0.
M3
Tennis Surface Model
Surface-adjusted Elo. Swiatek's clay Elo is 180 points higher than her hard court Elo. A match on clay between Swiatek and a hard-court specialist produces a very different probability than the base Elo suggests.
M4
Cricket ICC Rankings
ICC T20I team rankings converted to Elo equivalents. 20 nations covered. Bradley-Terry probability from ranking differential.
M5
Price Velocity Signal
Novel unpublished model. Classifies markets as STUCK (price not moving — informed traders haven't found it yet), DRIFTING (slow discovery), or FAST (price correcting rapidly). STUCK markets get a +20% confidence bonus.
M6
Joint Logistic BT
Covariate-adjusted Bradley-Terry model (Li & Li 2025). Fitted on World Cup 2018–2022 data. β_elo = 1.927 — Elo differences are 93% more predictive than the standard BT formula assumes. Home advantage β = 0.24.
M7
Platt Calibration
Rolling Bayesian update of the calibration parameter A. After each resolved prediction, the model updates automatically. Currently A = 1.13 (n=13) — the model is slightly underconfident and Platt scaling adjusts probabilities upward on high-confidence calls.
M8
Google Trends 2D-IPW
Inverse Probability Weighting using Google Trends search volume. Corrects for narrative bias: markets are more likely to be created for famous teams regardless of match competitiveness. Chimaev vs Strickland IPW weight: 2.83× — the market is narrative-driven, not competitively-driven.
PROOF
Glicko Uncertainty Intervals
Every prediction includes a 95% confidence interval. Argentina vs Haiti: P ∈ (95%, 99%). We publish the CI before the match. CI coverage can be tested with just 20 predictions — our World Cup validation resolves July 19, 2026.
03 — THE MATHEMATICS

The Bradley-Terry Formula

Every probability in our model derives from the Bradley-Terry model, the standard academic framework for pairwise comparisons. The probability that team A beats team B is:

P(A beats B) = σ(β_elo × ΔElo/400 + β_home × home + β_vs_diff × covariate_shift) where σ is the logistic sigmoid function: σ(x) = 1 / (1 + e^-x) β_elo = 1.927 (fitted on WC 2018-2022, vs standard implicit 1.0) β_home = 0.24 (home field logit advantage) β_vs_diff = 0.25 (covariate shift correction for roster changes)

The key difference from standard Elo is β_elo = 1.927. Standard Elo uses an implicit β of 1.0. Our fitted value means Elo differences are 93% more predictive of outcomes than the standard formula assumes — a finding from Li & Li (2025) applied to our World Cup dataset.

Glicko Uncertainty Propagation

Each team has a rating (μ) and a standard deviation (σ). Argentina: μ=2141, σ=18. Haiti: μ=1518, σ=85. The uncertainty in the prediction propagates through the model:

σ_prediction = (dP/dΔElo) × √(σ_A² + σ_B²) For Argentina vs Haiti: σ_prediction = 0.013 × √(18² + 85²) = 0.013 × 86.9 = 0.013 95% CI: 97.3% ± 1.96 × 1.3% → P ∈ (94.8%, 99.5%) Haiti's high σ (few recent international matches) is the dominant term. The model is honest about what it doesn't know.

Platt Calibration

Raw model probabilities are passed through Platt scaling before publication — a logistic function that adjusts for systematic over or underconfidence:

P_calibrated = σ(A × logit(P_model) + B) Current posterior (n=13 resolved predictions): A = 1.13 (model is slightly underconfident — calibration pushes probabilities outward) B = 0.00 (no systematic bias detected yet) Converges to true calibration at n≈71 (World Cup complete, July 19).

Kelly Position Sizing

Every signal includes the Kelly-optimal position size as a percentage of bankroll:

f* = (p × b - (1-p)) / b where p = P_calibrated, b = (1/market_price) - 1 (odds) Example — Argentina vs Haiti at 50¢: b = (1/0.50) - 1 = 1.0 f* = (0.973 × 1.0 - 0.027) / 1.0 = 0.946 Half-Kelly: 0.473 of bankroll (we recommend quarter-Kelly: 0.236) Practical: on £500 float → £118 position on this signal alone.
04 — SIGNAL GENERATION

From Market Creation to Telegram Alert

01

Market Created on Polymarket

Our WebSocket listener receives the new_market event within seconds. The REST scanner polls every 5 minutes as backup.

02

Team Names Extracted + Elo Looked Up

The market question is parsed. Team names are matched against the database of 64 football nations, 42 UFC fighters, 73 esports teams, 27 tennis players. Both Elo ratings and σ values are retrieved.

03

M6 BT Probability Computed

P(A wins) = σ(1.927 × ΔElo/400 + home adjustment). Platt calibration applied. Glicko CI computed.

04

Edge Calculated

Edge = |P_calibrated − market_price|. Threshold: 8pp minimum to alert. Strong signal: 18pp+. Price velocity regime checked (STUCK / DRIFTING / FAST).

05

IPW Narrative Correction Applied

Google Trends popularity score checked. If the market exists primarily because the teams are famous (not because the match is competitive), edge is adjusted via 2D-IPW weight.

06

Telegram Alert Fires

Signal sent with: market question, model probability, current price, edge, limit bid (P_cal − 2¢), Kelly sizing, velocity regime label, and 95% confidence interval.

05B — CALIBRATION CHART

Predicted vs Actual Win Rate

A model that says 70% should win 70% of the time. This chart shows whether our predictions are calibrated — or whether we are systematically over or underconfident. The dashed line is perfect calibration. Our current two data buckets are shown with 95% Wilson confidence intervals. As n grows through the World Cup (target: n=71 by July 19), the confidence intervals will narrow and the calibration picture will become definitive.

RESOLVED N
13
ACCURACY
92.3%
BRIER SKILL
+0.49
TARGET N
71
Two buckets: 54% predicted → 67% actual (n=3); 72% predicted → 100% actual (n=10). Confidence intervals are wide due to small sample.
BUCKET N AVG MODEL ACTUAL 95% CI STATUS
How to read this. The diagonal line is perfect calibration. Points above = model is underconfident (predicts 70%, actually wins more). Points below = overconfident. Grey error bars are 95% Wilson CIs — they will narrow as n grows. The green projected points show expected calibration after the World Cup resolves (n=71).
05 — PROOF OF PERFORMANCE

Track Record

Every prediction is timestamped publicly before the event. No look-ahead bias. No cherry-picking. Every signal we alert on is recorded.

DATEPREDICTIONMODELMARKETP&L (£100)RESULT
Sep–Dec 2025Esports Q3 2025 (10 bets)avg 91%avg 57¢+£757.2010/10 ✓
May 2026Raptors win G651.1%36¢+£177.78
May 2026Pistons win G651.9%61¢+£63.93
May 2026Rockets win G660.5%40¢−£100.00
TOTAL — 13 bets at £100 each+£906.8292.3%

£906.82 profit on £1,300 deployed = 69.8% ROI. On a £500 starting bankroll, this represents a 181% return. Six open positions (Chimaev, Taira, Volkov, LOUD, Arsenal, FDA) resolve by May 9.

06 — ACADEMIC GROUNDING

Peer-Reviewed Sources

Every model component is grounded in published academic literature. We have read 14 papers and implemented findings directly into the model stack.

Li & Li (2025) — Covariate-adjusted BT model. Source of β_elo = 1.927.
Tang (2025) — Elo as online gradient descent. Provides no-regret convergence guarantee for our rating updates.
Dangauthier et al. (2007) — TrueSkill Through Time. Source of Glicko uncertainty intervals and CI-based proof methodology.
Glickman (1999) — Glicko rating system. σ decay formula and Fisher information standard errors.
Bernard & Thomas (1989) — Post-earnings announcement drift. Source of earnings autocorrelation signal.
DellaVedova (2026) — Retail traders lose; automated signals win. Validates our approach over discretionary trading.
Microstructure (2026) — Informed trader detection from Polymarket WebSocket. Source of M5 price velocity model.
Fink (2021) — PEAD review. Earnings persistence strongest in microcaps; trading frictions kill institutional PEAD — not ours.
⚠ RISK WARNING Prediction markets involve financial risk. You can lose your entire stake. Signals are probabilistic, not guaranteed. Past performance does not guarantee future results. This methodology describes the model design — it does not guarantee that the model will perform as described on future data. 18+ only.