ogsfrompoly

Methodology

How ogsfrompoly separates skill from luck on Polymarket. How it scores past alerts. How it decides what is safe to publish.

ogsfrompoly asks a narrow question:

Which Polymarket wallets still look skilled once you strip the luck out — and do their moves keep working when we score them later?

We publish the answer after the fact, as a scorecard. We publish no live trade ideas, no open positions, no wallet leaderboards.

Current scope

The public track record covers one Polymarket area:

The narrowness is on purpose. We do not chase category coverage. We build a clean audit trail: skilled wallets, alerts scored after the fact, and outcomes a reader can check later.

How a wallet qualifies

A wallet needs enough resolved history to test. An event here means one real-world question that Polymarket settles. One question can carry many markets, outcomes, fills, or single token positions, and it still counts once. The bar is at least 20 events resolved inside the measurement period.

Only resolved markets count. We ignore open positions. For each eligible wallet we build an only-realized PnL series: one profit-and-loss number per resolved event.

We group at the event level because one prediction can span many fills or markets. Count each fill as its own decision and one position starts to look like a dozen independent bets. Grouping keeps the test on the thing we care about: did this wallet pick the right side, again and again, on questions that have already settled?

The skill test

We measure skill with an event-level sign-randomization test.

For each eligible wallet, the inputs are:

Each simulation keeps the size of every event result but randomizes the sign. For simulation j, each event gets a random sign:

s_{j,i} in {-1, +1}

The simulated no-skill total is:

T_j = sum(s_{j,i} * abs(x_i))

The p-value is:

p = count(T_j >= A) / B

In words, the p-value answers:

"If this wallet were just guessing, how often would random sign flips match or beat its actual total?"

We call a wallet skilled only when that p-value clears the pre-registered threshold: p <= 0.05. A big raw PnL number is not enough.

Controls before publication

The public scorecard measures repeatable trader signal. It must not measure address shuffling or artifacts of market microstructure.

Before a wallet reaches the published record, we screen it. We look for wallets moving as one coordinated pack, for the same counterparty on the other side again and again, and for trades that look like wash trades. A signal that fails gets cut or demoted.

How we score alerts

Weekly statements score the period after the fact. They report:

Monthly statements add the business side of the project: revenue, operating expense, net result, and runway.

The published category is macro-finance. Anything outside the public scope drops off the statement pages. We never merge it into another category.

Every statement looks back. We publish whether past signals worked. We publish nothing about a trade in flight.

Skin in the game: two small live agents

The scorecard looks back. The money does not. Two agents trade live behind our own signals, on tiny $5 tickets:

We document both in plain English at copy trading for dummies: every rule, every live setting, and a simulator you can drag. Their realized results reach the monthly statements as cash. We sweep profit to a collection wallet each week and count it as revenue once it lands. We publish nothing about an agent position while that position is open.

Academic foundation

The methodology rests on two 2026 papers about skill, trader persistence, and outcomes in prediction markets:

Both papers point at the same lesson. Short-term realized PnL is a weak proxy for skill. A wallet can get lucky. A lucky wallet can stop winning.

So ogsfrompoly starts with a statistical skill test. Then it publishes, in the open track record, whether those signals worked.

Disclosure policy

We are open about methodology and aggregate results. We are closed about anything that could front-run a position or expose a single trader.

These rules apply to the rendered HTML page, the raw .md feed, RSS, and llms.txt.

When a published number sits close to the line, we delay it 30+ days or anonymize the category. When in doubt, we don't publish.

What this is not

ogsfrompoly is not an investment recommendation, trading signal service, or claim that any wallet will keep winning.

It is an audit trail for one method. Find wallets whose resolved history does not look like random signs. Score their alerts after the fact, in the categories we publish. Show the results.