Free lesson · Prediction foundations

A causal forecast research ledger

Open interactive lessonPractice calculationsExplore labs

A valid forecast comparison uses only information available when the forecast was issued and outcomes that eventually matured.

Save event ID, contract text/version, decision timestamp, horizon, model probability, contemporaneous bid/ask, and resolution source. Keep an append-only original forecast even when later data corrections are necessary.

Use a chronological split by event, not trade row. Lock model and cost rules before the final evaluation. Report both forecast scores and separate implementable outcomes, including unfilled orders and unsettled events.

Apply it in a strategy

• Produce a forecast research memo with frozen chronology, matched baselines and a distinct trading-cost analysis.
• Record the input timestamp, executable quantity, currency and horizon. Reconcile the result with a cash-flow or state table.
• Stress this failure condition: Selecting domains or horizons after inspecting test scores turns the test set into a tuning set.

Research deliverable

Build and explain a a causal forecast research ledger worksheet. Produce a forecast research memo with frozen chronology, matched baselines and a distinct trading-cost analysis.

Evidence boundary: Synthetic arithmetic and scenarios illustrate mechanics. They are not historical returns, a paper replication, or evidence of an executable edge. Research sources and their access limitations are recorded at the end of this module.

Primary research and operational references · reviewed 12 September 2026

Further reading: Saguillo et al. — Unravelling the Probabilistic Forest ↗

arXiv preprint v1, 5 August 2025 · Review: Abstract, data-section excerpts and full-text structure inspected 2026-09-12. Markets/data: Polymarket markets resolved 1 April 2024–1 April 2025; data-section endpoint dates checked in the primary manuscript. Interpretation: Distinguishes within-market complete sets from cross-market logical relations. Limits: Historical detected opportunities depend on contract classification, timestamps and execution assumptions; not a current opportunity list. No independent replication performed.

Further reading: Le — Decomposing Crowd Wisdom ↗

arXiv preprint v2, revised 4 August 2026 · Review: Abstract, data-section excerpts and full-text structure inspected 2026-09-12. Markets/data: Kalshi and Polymarket; the primary data tables identify a 31 December 2025 cutoff; inspect exact extraction rules before reproducing. Interpretation: Motivates calibration by domain/horizon and explicit sampling uncertainty. Limits: Descriptive calibration and in-sample decompositions do not establish net trading profits; clustered events matter. No independent replication performed.

Further reading: Polymarket — How positions work ↗

Living official product documentation · Review: Documentation inspected 2026-09-12. Markets/data: Conditional outcome-token mechanics; no empirical sample. Interpretation: Operational starting point for split, merge and redemption workflows. Limits: Record the exact market rules and collateral/version before applying a generic binary model; platform behavior may change. No independent replication performed.

Research sources, review dates and limitations

Connect the ideas: Dependence and diversification

Retrieve: Joint behavior matters when combining uncertain outcomes.

Check the change: Correlation, cointegration, covariance and event dependence answer different questions.

Statistics → Linear algebra → Time series → Portfolio construction → Portfolio management theory

# Python 3.10+; standard library unless NumPy is imported below.
# Inputs and outputs use the units defined in this lesson. Synthetic teaching example.
def brier_skill(model_score,baseline_score):
    if model_score<0 or baseline_score<=0: raise ValueError("Nonnegative model and positive baseline scores required")
    return 1-model_score/baseline_score

print(brier_skill(.18,.24))