Testing prediction-market signals with point-in-time provenance

Prediction-market odds can offer a second view of an event that matters to a trade. The challenge is establishing when that view was available and what the contract actually resolved. DXAP currently describes Polymarket Alpha as odds alongside price. We describe a PROPOSED experiment to test the information increment of that input under a fixed forecasting and trading task. This is an evaluation design, rather than a finding that odds improve returns.

Save the contract, then save the clock

For every input, we would retain the contract identifier, resolution terms, quoted probability, source timestamp and local receipt time. The trading decision time determines eligibility. If a quote arrived after that time, it belongs outside the model's information set even when its source timestamp appears earlier. Historical corrections and revised records require versioned provenance so later data repairs cannot enter an earlier replay silently.

The contract's proposition should map to a specific market hypothesis. An election outcome and an asset's next-hour return are different events. We would write the hypothesized transmission mechanism before fitting a model: for example, an odds change may alter expectations about policy relevant to the asset. That mechanism defines the lag windows to test and the circumstances in which the feature should be irrelevant.

A fixture for a misleading timestamp

In an illustrative case, a quote bears a 12:00:00 timestamp, reaches our collector at 12:00:08 and the agent decides at 12:00:05. A replay using only source time grants the agent three seconds of unavailable information. The eligible snapshot should instead use the latest quote received before 12:00:05. We would include this case in the data acceptance tests and publish how missing or late quotes affect coverage.

Liquidity also changes interpretation. A stale last trade is a different measurement from an executable bid and ask. We would preserve spread and quote age, then compare informative liquid contracts with low-activity contracts under declared categories. Treating all probabilities as equally fresh can turn a timing artifact into apparent predictive skill.

Compare one additional input at a time

The reference arm receives market and portfolio state. The treatment receives that same snapshot plus eligible prediction-market fields. Model, question, risk policy and evaluation horizon stay fixed. We would score forecasts first, then simulate decisions under explicit costs. An improvement in explanation quality should be reported separately from an improvement in probability scores or net results.

Our continuous record includes a retracted lag-edge claim caused by a timestamp artifact, making clock discipline a concrete lesson from the research lineage. The controls paper explains why complete rendered inputs belong in each trace. A useful next step for agentic trading is therefore an inspectable point-in-time source record: it lets us test when an external signal helps, where it fails and whether the apparent increment survives a genuinely later evaluation window.