Does the route to halftime matter, or is the halftime market price enough?

The study asks whether the sequence of first-half market updates adds predictive value beyond the pre-match price P₀. A secondary question compares explicit path-based features with the halftime price level P_HT, which may already summarize the market's accumulated first-half information.

Interpretation

The results suggest that Polymarket prices behave like probability-like signals that update as match information arrives—goals, momentum shifts, expected-goals narratives, red cards, and tactical adjustments all tend to move prices toward contracts that better reflect the evolving match state.

A useful economic intuition is that by halftime, market participants have often already incorporated most of the first-half news into the level P_HT. Explicit path-based features may therefore repeat information that is already absorbed into that endpoint: under simple linear models with four intra-half snapshots, recording where the price lands at halftime can act as a parsimonious summary of first-half updating.

For parsimonious forecasting under the current data and model class, storing P_HT may be sufficient relative to engineering finer-grained path coordinates. More complex path modeling may become valuable once richer microstructure data are available (higher-frequency quotes, order-flow, liquidity, or trading volumes).

These conclusions describe predictive associations under the stated evaluation protocol; they are not causal claims about which features “matter” in an economic sense—only about what forecasts best under this sample and model family.

Limitations

The findings should be interpreted conservatively. The sample covers 309 football events (927 contract-side conditions); performance gains are heterogeneous across grouped chronological folds (mean CV log-loss improvements versus P₀ vary roughly 0.02–0.08 units across folds).

The extract does not include liquidity, trading volume, or order-flow data, so the study cannot distinguish a large price move on thin volume from the same move supported by deeper market activity—both appear as identical price updates in the feature set.

The evaluation uses logistic regression on raw bounded features; richer nonlinear or tree-based specifications were out of scope and might interact differently with path coordinates. The held-out test split can disagree slightly with cross-validated rankings for nearby models (for example, reported test log losses for some path specifications versus the halftime benchmark), underscoring single-split uncertainty.

Contracts follow a three-way structure with a positive-outcome rate near one-third; results may not transfer mechanically to other sports, leagues, or binary-only markets.

Future Extensions

Higher-frequency snapshots (for example one-minute marks or order-book midquotes) could test whether finer path detail recovers incremental signal once the endpoint alone stops being a sufficient statistic.

Liquidity- and volume-weighted price changes would treat a 10-cent move on thin depth differently from the same move with sustained participation—directly addressing the largest blind spot in the current feature set.

Nonlinear models (gradient boosting, shallow neural nets) could capture interactions between interval moves and baseline levels that logistic regression cannot express cleanly.

Larger multi-season and multi-league panels would stabilise fold-level estimates and support cross-sport generalisation—for instance whether a halftime-level-dominates-path pattern holds in basketball, tennis, or other timed formats.

Finally, combining path features with non-market signals (expected-goals models, team-level priors, injury news) would clarify whether market path information remains incremental once conditioning on observable fundamentals outside the order book.