DV-PMI · Insider screen
Private Information Index (PII)
Across 4,657,827 testable trader-event pairs (1,020,455 traders, 27,361 multi-market events), the per-event joint-accuracy test produces 1,008 raw flags. Excluding mutually exclusive events and dependent bet structures leaves a distinct-question core of 112 trader-event pairs, of which 11 survive the within-event dependence adjustment at the five percent level and 0 at one percent. On a placebo class where episodic private information is implausible, the corrected test yields zero discoveries.
The detection funnel. Each stage removes a source of false positives: platform-defined mutually exclusive (neg-risk) events, dependent bet structures (single-game multi-bets, cumulative threshold ladders), and residual within-event correlation (beta-binomial adjustment at intraclass correlation 0.18, estimated from the data).
What this index measures
Detecting informed trading requires separating transient information from persistent skill, and the separation cannot be made at the trader level: pooled averages dilute episodic information, and the best-episode statistic inflates mechanically with activity. The test therefore operates at the trader-event unit. For each trader and each multi-market event where the trader took positions in at least two component markets, the joint probability of the observed directional accuracy is evaluated against the price-implied null: a trader with no information does no better than the price she paid. Multiplicity is controlled across the events each trader contests (Holm within trader).
Survivors by disclosure archetype
| Archetype | Pairs |
|---|---|
| speech-content | 7 |
| document-or-decision | 3 |
| enumeration | 1 |
| Total (5% level) | 11 |
All survivors are all-correct records on disclosure events (adjusted p-values 0.031 to 0.048). They are reported by archetype only; the dashboard names no wallets. Flags mark statistical patterns consistent with informed trading, not legal determinations. Net directional profit across the 74 core pairs with position coverage is $501,167 (median $778 per pair).
Placebo calibration
The false-positive rate is calibrated on a real-data placebo: 597,146 testable pairs in asset-price-direction markets, where episodic private information about the fundamental is implausible. The corrected test yields 0 discoveries on the placebo at every threshold examined. An uncorrected best-episode rule, by contrast, flags active traders at a rate that rises mechanically with how many events they trade.
Forward test
The statistic flags the trader charged in CFTC v. Spagnuolo at the five percent level under the estimated within-event dependence; it was first run sixteen days before the complaint. Flags mark statistical patterns consistent with informed trading, not legal determinations. The wallet mapping is the author's inference from public on-chain data, not a fact stated in the complaint.
Sustained-skill monitor (reclassified wallet-level screen)
This wallet-level pooled excess-accuracy screen was the April 2026 draft's detector. The June 2026 revision reclassifies it: a statistic that averages a trader's record measures sustained skill and is asymptotically blind to episodic information. It is published here as a skill-persistence monitor, not an informed-trading detector.
| Wallet class | Wallets tested | Flagged (p < 0.01) | Flag rate |
|---|---|---|---|
| Algorithmic | 143,084 | 2,832 | 1.98% |
| Sophisticated | 92,713 | 1,620 | 1.75% |
| Active Retail | 470,086 | 6,033 | 1.28% |
| Total | 705,883 | 10,485 | 1.49% |
Of the 10,485 wallets with sustained excess accuracy at the one-percent threshold, 1,558 survive the Holm-Bonferroni family-wise correction and 4,700 survive the Benjamini-Hochberg FDR procedure. July 2026 rerun on the corrected master (trades through 2026-07-03; wallets with at least 10 resolved trades).
Methodology
Per-event joint-accuracy test against the price-implied null (each position's success probability is its own transaction-time price), Holm-corrected across the events each trader contests, mutually-exclusive (neg-risk) events excluded, and a beta-binomial dependence adjustment at intraclass correlation rho = 0.18. Event groupings and the mutually-exclusive attribute come from the platform's own market metadata; disclosure archetypes are assigned by a rules-and-language-model classification validated against hand-coded markets. Full construction, proofs, and robustness are in the source paper.
Source paper
Della Vedova, J. (2026). Detecting Informed Trading in Prediction Markets: One Event at a Time. SSRN 6567238, revised June 2026. The companion paper, Picking Winners and Losing Money: The Two Dimensions of Trading Skill, establishes the accuracy-execution orthogonality that this paper inverts into an identification problem.
Related surveillance indices
PII is the first of eight DV-PMI surveillance indices. The companion indices test for pre-resolution timing concentration, spread patterns in markets with flagged wallets, and the response of flagged-wallet excess accuracy to resolution surprises. See the full surveillance suite →
Data
pii_snapshot.csv · pii_latest.json
Cite this index
@misc{dellavedova2026pii,
title = {Private Information Index},
author = {Della Vedova, Joshua},
year = {2026},
url = {https://jdellavedova.com/pii}
}