Signal methodology

How It Works

The methodology behind every signal — no black boxes.

01

How We Identify Best-Performing Wallets

Not every active wallet belongs in the signal pool. We apply four hard filters before a wallet is tracked.

  • Statistical Confidence

    Each wallet must clear a Bayesian credible interval test in its primary market category. We use Beta-Binomial empirical Bayes with category-specific priors fit from that category's historical base rate — a wallet at 65% with 8 resolved bets looks very different from one at 65% with 80 resolved bets. The posterior 90% credible interval must exceed the category's fee-adjusted breakeven rate.

  • Position-Weighted Volume Floor

    Wallets must have transacted at least $50,000 of resolved-position notional (raised to $100,000 for Mentions markets). Volume is more skill-correlated and more sybil-resistant than trade count — faking 500 trades costs pennies; faking $50K of resolved notional costs $50K.

  • Minimum Resolved Bets

    At least 10 confirmed resolved bets are required (15 for Mentions, where intra-event correlation reduces effective trial count). This is not a substitution for the statistical test — it ensures the posterior has meaningful data.

  • Market maker filter

    Wallets that predominantly provide liquidity — hedging both sides across many markets rather than expressing directional conviction — are excluded entirely. Their positions would produce false signals.

The wallet pool is re-scored weekly via an automated pipeline. Wallets that fall below the thresholds are removed; newly qualifying wallets are added.

See per-category methodology →


02

How The Confluence Score Works (0–10)

Each signal is scored across five independent layers, listed below in order of influence. Internally the layers combine into a composite, which we publish rescaled to a 0–10 scale (one decimal, e.g. 8.4) — the same ARCANIQ Score shown on the wallet board.

Wallet rank scoreLargest factor

Which wallets are betting this side, weighted by their ARCANIQ Score — realized risk-adjusted return, entry timing, consistency, and verified sample size. (We rank on return, not how often a wallet is right.)

Early mover signalMajor factor

Are they entering before odds shift? Measures timing edge relative to the broader market.

Liquidity impactMajor factor

Bet size relative to order book depth — a proxy for conviction.

MomentumSupporting factor

Odds movement velocity — how quickly the market is repricing in the signal direction.

Volume surgeSupporting factor

Unusual volume relative to baseline market liquidity. Normalised to reduce noise.

Dynamic weight adjustment: Weights are dynamically adjusted based on signal correlation — correlated signals have their weight reduced to prevent double-counting.


03

What Closing Line Value (CLV) Means

Hit rate tells you how often you were right. CLV tells you if you had edge.

A market can resolve in your favour because of luck. CLV measures whether you entered at better odds than the market's final assessment before resolution — a more reliable indicator of genuine edge.

Example

A signal entering YES at 34¢ that closed at 60¢ before resolving YES = strong edge (+26¢ CLV)

We track CLV on every signal. Positive average CLV means our signals beat the market's final assessment, not just the resolution.


04

How We Prevent Lookahead Bias

Lookahead bias is the most common way a backtested signal looks good on paper but fails in production. We address it at every layer.

  • public_date, not event_date

    All signals use the date the market became publicly visible, not the underlying event date. This mirrors what was knowable in real time.

  • Out-of-sample validation (chronological 80/20)

    Weights are optimised on the first 80% of historical data, then evaluated on the held-out final 20%. The split is chronological, so the model never sees future data during training.

  • Out-of-sample degradation warning

    If test-set accuracy drops more than 5 percentage points below training-set accuracy, an automatic warning is raised and the weight set is reviewed before deployment.


05

How We Handle Losing Signals

Cherry-picking wins while burying losses is the oldest trick in financial services. We don't do it.

  • Every loss is posted publicly

    Losing signals are displayed in the same format as wins — same data, same visibility. The full signal history is publicly accessible.

  • Automated loss explanation within 24 hours

    When a signal resolves against the position, an automated email is sent to all subscribers within 24 hours explaining the resolution — what the signal was, what happened, and what the CLV was.

  • No cherry-picking, ever

    The public signal log is append-only. Nothing is removed. Accuracy statistics are computed over the complete record, including all losses.


06

What EV Means

EV (expected value) is the difference between our estimated probability and the market's implied probability, after fees.

Our estimated probability
Market implied probabilityP_market
Executable fee + spreadf

EV = P̂ − P_market − f

f is the executable cost of the trade: Polymarket's taker fee (category-based and price-dependent — it peaks near 50¢ and makers pay none) plus the order-book spread. We subtract the actual applicable cost, not a flat assumption.

  • Positive EV = mispriced market

    Positive EV means our model estimates the market is pricing an outcome lower than the evidence supports, after accounting for fees.

  • We only alert on positive EV signals

    No signal fires unless EV is positive. The size of EV is reflected in the signal label — a higher EV score gets a higher confluence weight.

Example label

EV: +18.4% · EDGE

Our model estimates the true probability is 18.4 points above the market's current price, net of fees — a positive expected value. Past accuracy does not guarantee future results.


07

Suppression Transparency

Hit rate shown elsewhere is computed on delivered alerts — the signals that passed our filters. We don't claim that delivered subset is an unbiased estimate of every candidate signal; instead we publish the full suppression volume below so you can see the size of the funnel behind that number, and we report the complete delivered record with no cherry-picking (see §05). Every scoring pass scans the active market universe; the bulk are filtered out before any subscriber is notified.

  • Negative-EV filter

    Any signal whose lower-bound EV after fees and the executable spread is below zero is dropped. Subscribers never see a NEGATIVE EV alert.

  • Low-conviction noise filter

    Signals in the low-conviction band — where the position size implied by our edge is negligible and statistical confidence is too low to act on — are suppressed regardless of the score headline.

Last 7 days suppressed

3,485,339

0 negative-EV

3,485,339 low-conviction noise

Updated every 5 minutes. The number reflects candidate signals dropped pre-delivery — the funnel size behind the hit rate you see published. Past accuracy does not guarantee future results.


This platform is for educational purposes only. Not investment advice. Past signal accuracy does not guarantee future results. All prediction market participation carries risk.