Prediction
markets
superstudy.

• Translate a synthetic binary contract into a cash-flow table.
• Distinguish a probability estimate, a quoted price, and a realized outcome.
• Separate event contracts from perpetual derivatives and separate access from permission.

Begin with the payoff, not the headline. In this course a synthetic YES unit costs a stated amount now and pays $1 if a precisely defined event resolves YES, or $0 if it resolves NO. This is a teaching convention; real contracts can have different settlement rules, fees, cancellations, and remedies. Owning the YES side does not guarantee that the popular interpretation of a headline will be the event that actually settles.

A price of $0.60 can be a useful starting point for discussing a 60% probability, but it is not proof of that probability. Fees, spreads, risk preferences, position constraints, limited participation, and scarce liquidity can all create a gap between a quote and an aggregate belief. A last trade is historical; an ask is a price someone is offering; your own probability is a forecast that can be wrong. These three objects should have different labels in a paper journal.

An event contract resolves under a rulebook. A perpetual derivative generally tracks an underlying exposure without a fixed expiry and can involve margin, funding, and forced liquidation. Neither an advertisement nor an API endpoint establishes that a particular person in California can access a particular product. Keep four independent checkboxes: what the product is, what the contract permits, what current law and regulator records say, and whether the account holder meets current eligibility requirements.

Basics of Futures TradingPredictions and Perps API documentationWhat is Polymarket US?Dated primary source shelf

• Choose a relevant reference class before reacting to a vivid signal.
• Distinguish P(evidence | event) from P(event | evidence).
• Update a prior using explicit likelihood assumptions and test sensitivity.

A base rate is the frequency of an event in a relevant reference class. The difficult part is choosing a class that is comparable to the present situation: similar definitions, incentives, observation windows, and data quality. A striking anecdote does not erase a base rate. Begin a forecast with the class, the sample size, the uncertainty, and the reasons the current case might differ.

Bayes' rule is P(H | E) = P(H)P(E | H) / [P(H)P(E | H) + (1 - P(H))P(E | not H)]. The probability that a signal appears when an event is true is not the probability that the event is true after seeing the signal. Both the true-signal rate and the false-signal rate matter. If the evidence has zero probability under both possibilities, the supplied model cannot update; the posterior is undefined.

Likelihood inputs are often estimates rather than measured constants. Report a range when the signal quality is uncertain, and avoid counting the same evidence twice. Two articles repeating one source are not two independent confirmations. When evidence arrives over time, document the information available at each update so a later result cannot silently rewrite the earlier forecast.

Dated primary source shelf

• Score recorded binary forecasts without rewriting them after settlement.
• Distinguish calibration from discrimination and profitability.
• Recognize sample size, selection, and dependence limits in a forecast record.

A calibrated forecaster's 70% forecasts should resolve YES roughly 70% of the time across a sufficiently informative collection of comparable predictions. Calibration is a property of a collection, not an individual event. A 90% forecast that resolves NO is not automatically a bad forecast; persistent underestimation of the remaining 10% is. Keep time-stamped forecasts, definitions, revisions, and outcomes before calculating scores.

For binary outcomes y in {0, 1}, the Brier score is the mean of (p - y)². Lower is better, with 0 for perfect predictions and 1 for completely confident wrong predictions. Honest probabilities minimize expected Brier loss under the forecaster's belief. That does not mean a tiny observed score difference proves skill. Compare with a relevant baseline, such as an available base-rate forecast made with the same information, and study uncertainty rather than just a leaderboard.

Calibration asks whether stated frequencies match observed frequencies. Discrimination asks whether higher forecasts tend to identify the events that happen. A forecaster can be calibrated yet uninformative by assigning a suitable common base rate to everything. Profitability is a further question involving the available price, fees, execution, and size. Do not infer trading ability from a quiz score or a small, selected paper sample.

Strictly Proper Scoring Rules, Prediction, and EstimationDated primary source shelf

• Extract the controlling question, deadline, timezone, and resolution source.
• Distinguish real-world ambiguity from a contract's specified settlement procedure.
• Model oracle, correction, cancellation, and dispute risk.

Your model must predict the event defined in the contract. Read the complete question and rulebook: what observation counts, which source controls, what time and timezone close the window, whether a first release or a revised value matters, and whether rounding changes the threshold. A headline such as 'above 100' leaves unanswered whether exactly 100 qualifies, which measurement is used, and whether a correction published later can change settlement.

An oracle is a mechanism that brings external facts into a settlement process. It might be an exchange's determination under its rules or a decentralized proposal and challenge procedure. Neither design removes the need to inspect who supplies evidence, who can object, what incentives apply, how long disputes take, and what exceptional outcomes are possible. The course does not assume that one venue's oracle or appeal rights apply to another.

Settlement risk belongs in the analysis even when your real-world forecast is excellent. Ambiguous wording, source outages, delayed publication, revisions, canceled events, and disputes can alter the payout or tie up paper capital longer than expected. A course summary cannot replace the controlling rulebook. Preserve its dated version and record uncertainties before assigning a probability.

Polymarket US RulebookDated primary source shelf

• Distinguish the best bid, best ask, last trade, and available depth.
• Calculate fee-adjusted expected value and weighted price across a synthetic ask book.
• Recognize partial fills and why a static paper fill overstates execution certainty.

The bid is a quoted buying price and the ask is a quoted selling price. Their difference is the spread. Depth is the quantity offered at each price, not a promise that the quantity survives until an order arrives. A last-trade chart can be stale, while a midprice may lie between quotes where no immediate transaction is offered. Keep price freshness, bid and ask sides, and quantities explicit in every exercise.

For a synthetic YES unit paying $1 or $0, with your probability p, purchase price c, and fixed fee f paid per unit regardless of outcome, expected net per unit is p - c - f. For q units, multiply by q. Break-even probability is c + f, which can exceed 100%; do not clamp it into an apparently attainable forecast. Real fee schedules may depend on price, outcome, order type, or other terms and need their own verified calculation.

Walking an ask book increases the average cost when the cheapest depth is exhausted. Fees, queue position, cancellations, latency, adverse selection, and market movement create further differences between a displayed opportunity and a fill. A limit order can remain unfilled; an immediate order can use only the depth that actually exists. The calculator uses a frozen synthetic book, returns incomplete fills honestly, and cannot estimate live execution quality.

Dated primary source shelf

• Distinguish independent losses from several positions sharing one underlying risk.
• Calculate compounded drawdown and the gain needed to recover.
• Understand the analytical Kelly formula and why its assumptions are fragile.

Ten contracts are not ten independent risks if they all depend on one election, announcement, data release, or settlement source. Positive correlation can make losses cluster. Capital may also remain unavailable while events settle or disputes resolve. List exposure by underlying driver and by shared infrastructure failure, then stress simultaneous losses. Counting contract names is not a diversification analysis.

If a paper portfolio loses fraction r of its remaining capital on each of n consecutive losing exercises, the drawdown is 1 - (1 - r)^n. At r = 10% and n = 5, the loss is 40.951%, not 50%; the remaining capital is 59.049% of the start. Recovering a drawdown d requires a gain of d / (1 - d) on the remaining capital. Losing half requires doubling what remains. This arithmetic does not estimate the probability of a losing streak or guarantee that ruin is avoided.

For one idealized binary wager with total cost c per $1 unit, 0 < c < 1, probability p, and net win odds b = (1 - c) / c, expected log wealth is p ln(1 + bf) + (1 - p) ln(1 - f), where f is the fraction of capital exposed to a complete loss. The unconstrained Kelly optimum is f* = (pb - (1 - p)) / b = (p - c) / (1 - c). This is an analytical result under known probabilities, repeatability, and the assumed payoff and cost. It is not a recommended position size. Estimation error, correlated exposures, changing opportunities, capital locks, execution limits, and different preferences undermine those assumptions; a negative formula does not authorize shorting.

Dated primary source shelf

• Write a falsifiable hypothesis with an explicit information timestamp.
• Detect look-ahead leakage, selection bias, and repeated testing.
• Evaluate forecast quality and paper execution separately.

A hypothesis specifies why an observable signal could predict a defined event better than a relevant baseline. Record the signal, available timestamp, forecast rule, comparison, and criteria that would count against the idea before examining outcomes. 'Buy things that went up' is not a complete hypothesis; the information set, timing, costs, and failure conditions are missing. Make the initial research claim small enough to be disproved.

Look-ahead leakage occurs when a historical decision uses information that was unavailable at the decision time, including later revisions, final settlement text, future categories, or cleaned data published afterward. Survivorship and selection bias occur when vanished, canceled, illiquid, or losing markets disappear from the sample. Trying many variations and reporting only the winner turns random noise into apparent skill. Keep a record of all tested ideas and reserve genuinely untouched observations.

Split evaluation forward in time and separate related events across training and evaluation when necessary. Fit transformations on training data only. Use only data whose collection and research uses are permitted; a visible quote or downloadable API field is not an unrestricted license. Report assumptions about missing observations, executable prices, costs, partial fills, and capital locks. A paper backtest can support a research hypothesis under those assumptions, but cannot reproduce queue position, market impact, or actual obligations.

Dated primary source shelf

• Recognize manipulative activity and avoid making it a strategy hypothesis.
• Treat nonpublic information, event influence, and role-specific restrictions as review triggers.
• Separate public visibility, API access, redistribution rights, and AI-use permissions.

Market integrity is a prerequisite for learning from prices. Spoofed displayed interest, wash trading, misleading statements, coordinated distortion, and interference with an event or resolution process can damage that information. An apparent price anomaly is not permission to create one. This course studies how suspicious behavior degrades evidence and execution; it does not provide tactics for manipulation or evasion.

Inside information and conflicts require particular care in event markets. Rules may restrict people with material nonpublic information, influence over the event or settlement, or specified roles such as officials and participants. Applicable law and each venue's current rules need separate review. Do not infer that everything is permitted merely because a restriction familiar from securities markets has a different name or scope here. A course cannot provide legal clearance for a personal situation.

Data rights are equally specific. Permission to call an API can differ from permission to retain quotes, share them with third parties, republish them, train a model, or feed them into an AI service. Terms can impose attribution, retention, rate, commercial-use, and AI-use restrictions. This public course uses invented inputs; it must not ingest licensed venue quotes into an LLM or publish them without verified permission. Public factual descriptions can be sourced separately without building a redistributed market feed.

Source agency and trading prohibitionsPolymarket US market integrity policyCFTC Enforcement Division prediction-markets advisoryDated primary source shelf

• Explain a simulated order lifecycle and why intent differs from a confirmed fill.
• Design validation, idempotency, reconciliation, and failure handling conceptually.
• Identify the gap between a paper simulator and a live execution system.

A rigorous paper system separates permitted input data, a versioned forecast, a proposed paper action, pre-action checks, simulated matching, a position ledger, and reconciliation. Keep forecasts and execution assumptions distinct. A read-only concept diagram can show where a permitted public fact might enter, but this course does not create credentials, attach accounts, connect a wallet, or provide a live-order connector. No helper here can reach a venue.

Model order states explicitly: proposed, rejected, acknowledged, partially filled, filled, canceled, and unresolved. A request timeout means the status is unknown, not necessarily that nothing happened. Idempotency and durable intent identifiers prevent a retry from becoming a duplicate action; reconciliation compares recorded intentions, reported fills, positions, and cash. Out-of-order messages, stale snapshots, duplicate events, and cancel-fill races belong in a simulated failure exercise before any operational claim.

A static simulator cannot recreate queue priority, hidden liquidity, adverse selection, latency, venue downtime, market impact, real capital constraints, or legal and contractual obligations. Advanced study means making these limits explicit and testing accounting invariants, not declaring paper gains achievable. Conceptual controls include bounded inputs, permission checks, stale-data handling, audit trails, human review, and stopping when state cannot be reconciled. Their design is not authorization for autonomous trading.

Alpaca paper trading documentationDated primary source shelf

• Distinguish leveraged perpetual exposure from a bounded-payoff event unit.
• Calculate synthetic long equity, funding costs, and a terminal maintenance test.
• Explain path-dependent liquidation and how losses may exceed posted margin.

A perpetual derivative generally has no fixed expiry and uses mechanisms such as funding to support its relationship with an underlying reference. Funding may be paid or received and can change over time. Initial margin is collateral against a larger notional position; it is not the full size of the exposure. Mark-price rules, index composition, maintenance tiers, liquidation procedures, fees, and deficit treatment must be read for the exact product. An advertisement showing leverage proves none of those details or an individual's eligibility.

The synthetic calculator models one linear long held to a terminal observation. Initial notional N = collateral C × leverage L. For an underlying return m, P&L = N × m. Constant funding is N × (funding basis points / 10,000) × periods, charged on initial notional. Terminal equity is C + P&L - funding; terminal maintenance is ending notional × maintenance percentage. Equity at or below that threshold is flagged as a breach in this invented model. It does not calculate a real liquidation price, simulate early closing, or use any venue's margin engine.

Linear price P&L does not make leveraged risk simple. Funding, changing maintenance, gaps, forced liquidation, fees, and the inability to survive an adverse path can create discontinuous and path-dependent losses. A position that ends with positive equity may already have crossed maintenance and been closed earlier. During a gap or a failed close, losses can exceed posted collateral and an account may owe a deficit depending on the product and its rules. This chapter is a stress exercise, not evidence that the user has opted into or can access live perpetuals.

How Funding WorksHow Margin WorksPerpetual futures, explainedDated primary source shelf

• Compare products and venues with dated primary evidence rather than brand associations.
• Keep US and California access, global products, API approval, and data rights separate.
• Identify which assertions remain unresolved and need a fresh official check.

A current venue comparison is an evidence matrix, not a ranking of where to trade. Each row needs the exact legal entity and product, contract type, governing rules, settlement process, geographic and individual eligibility, fee basis, API availability, production approval requirements, and data-use permissions. Date the evidence and distinguish a regulator record, binding terms, developer documentation, marketing, and your inference. An official source can answer only the questions it actually covers.

Treat global Polymarket and Polymarket US as separate product and eligibility reviews; one name does not transfer access conditions between them. For Kalshi, review the current contract rules, account requirements, API and data terms, regulator materials, and California-specific product disputes in the dated source ledger. A sports-contract legal proceeding does not automatically decide every non-sports contract, API permission, or person's access. Where production permission or AI/data rights are unresolved, keep the claim unresolved rather than replacing it with a yes.

Adjacent services belong in clearly labeled comparison rows. A crypto-routing API such as Jupiter raises software-license and data-use questions that are different from a blanket conclusion about every spot transaction. A paper environment such as Alpaca's is an infrastructure comparison, not evidence that it hosts the same event contracts. Recheck product advertising against official product pages and regulator records. A claim about the 'first US perpetuals exchange' needs an exact product definition, date, relevant authorization, and access terms before it becomes course fact.

Dated primary source shelf

• Describe conditional 2027 scenarios without presenting them as forecasts or facts.
• Assign observable triggers, disconfirming evidence, and update dates to each scenario.
• Integrate probability, settlement, execution, integrity, and permission analysis in a paper capstone.

Scenario 1: Broader regulated distribution. If more exchanges and brokers expose event contracts in 2027, access channels and product differences may become more important to study. Observable triggers: New CFTC registrations or designation amendments; Official broker and exchange notices; API releases and documented onboarding changes. Uncertainty: Hypothetical scenario, not a forecast. No probability assigned; it can coexist with restrictions and does not imply personal eligibility.

Scenario 2: Access fragments by state or contract category. If disputes and rule changes persist, 2027 access may vary more by state, sports versus other events, or user class. Observable triggers: Published court orders; State regulator notices; Venue rule changes or contract withdrawals; Current geolocation-policy changes. Uncertainty: No legal outcome is assumed. An allegation is not a judgment; a case in one state does not establish another state's rules.

Scenario 3: Data licensing becomes a larger constraint. If venues clarify or tighten redistribution, derived-data, retention, and AI terms, a technically working API may still be unsuitable for a public learning product. Observable triggers: Revised data terms; Published Market Data Agreements; Express AI or retention clauses; Written licensing permission for the intended use. Uncertainty: Future licensing terms are unknown. No permission for this course to ingest or redistribute live venue data has been obtained.

Scenario 4: Stronger integrity controls. If regulators and venues add controls in 2027, confidentiality duties, outcome influence, contract exclusions, and surveillance evidence may receive more attention. Observable triggers: Updated rulebooks; Regulator enforcement notices; Exchange disciplinary publications; New market-specific exclusions or government ethics rules. Uncertainty: The scope, timing, and effectiveness of future controls are unknown. Enforcement allegations must be labeled accurately.

Scenario 5: Liquidity changes unevenly. If activity expands or access contracts in 2027, popular markets may gain depth while obscure or disputed contracts become harder to exit. Observable triggers: Authorized observations of executable spread and depth; Fee-schedule changes; Market-maker notices; Trading suspensions or settlement delays. Uncertainty: No growth or deterioration is predicted. Volume and publicity are not proof of executable liquidity, and paper fills do not establish live performance.

Scenario 6: Perpetuals expand, or margin rules tighten. Additional asset-specific approvals and product launches could broaden perpetual offerings in 2027. Volatility events could also lead to stricter margin, access or liquidation rules; both paths can coexist. Observable triggers: New CFTC product approvals; filed contract specifications; official changes to funding, collateral, maintenance and access requirements. Uncertainty: Neither new assets nor stable leverage limits are assumed. Funding, deficit protections and retail availability may differ by account and product.

These six conditional cases may coexist. Check effective dates, appeals, source rights and the exact product and jurisdiction. None is a factual claim about what 2027 will bring, a price forecast, or a recommendation to deploy capital. A developer endpoint alone is weak evidence for research permission; a press release alone cannot establish a final operative legal outcome.

Dated primary source shelf