BirdBets: A Trust-Minimized Parimutuel Prediction Market for Avian Feeder-Visitation Forecasting on a Layer-2 Settlement Network

Prediction markets aggregate dispersed private beliefs into a single, continuously-updated price that approximates the probability of a future event [1, 2]. Their canonical subjects are elections, sporting outcomes, and macroeconomic releases. We observe that the literature has, to date, almost entirely neglected the question of how many birds will show up.

This omission is curious. Bird-feeder visitation is a high-frequency, locally-observable, and emotionally resonant random process. It produces a fresh, verifiable data point every single day. It is, in the language of market design, an excellent settling instrument: cheap to observe, hard to fake at scale, and — crucially — extremely fun to argue about.

BirdBets fills this gap. The contribution of this work is threefold. (i) We specify a minimal parimutuel contract that resolves daily markets against a physical sensor. (ii) We describe an autonomous oracle pipeline that creates, seeds, resolves, and settles markets without human intervention. (iii) We provide the first rigorous treatment, to our knowledge, of the seed-supply attack — wherein a motivated actor manipulates the settlement value by feeding the birds — and argue that it is a feature.

Automated market makers for prediction. Hanson’s logarithmic market scoring rule (LMSR) [1] provides bounded-loss liquidity for thin markets and underpins much of the modern onchain prediction stack. Order-book and AMM hybrids such as Polymarket and the Augur protocol [2] extend this to permissionless settlement. BirdBets deliberately forgoes a scoring rule in favor of the older and simpler parimutuel design (§4), trading marginal price-discovery efficiency for a contract that a single reader can verify in an afternoon.

Oracles. The oracle problem — importing off-chain truth into a deterministic state machine — is the central difficulty of any real-world settlement system [3]. Decentralized oracle networks distribute this trust across many reporters. BirdBets, by contrast, adopts a frankly monarchical oracle: one address, one feeder, one source of avian truth. We justify this choice in §5 and §8.

Ornithology. The biological literature confirms that birds visit feeders, that they do so more on some days than others, and that the resulting time series exhibits weather-, season-, and seed-dependent structure. We cite this body of work collectively and gratefully [4] and decline to reproduce it.

The protocol comprises four loosely-coupled components, connected in a single daily loop:

1. The feeder. A BirdBuddy smart feeder observes the physical world and emits visit events.
2. The counter. A Home Assistant counter.bird_buddy_daily_visits entity accumulates those events into a single integer per local day, exposed over an authenticated REST API.
3. The oracle. A server-side automation reads the counter and submits transactions to create, seed, resolve, and settle markets.
4. The settlement contract. An ERC-20 parimutuel market on Base holds stakes in escrow and distributes payouts on resolution.

  day d-1 close          day d open            day d close          day d+1
       │                     │                      │                    │
  ┌────┴─────┐         ┌─────┴─────┐          ┌─────┴─────┐        ┌──────┴──────┐
  │ createMarket(d)    │  betYes / │          │ betting   │        │ resolveMarket
  │ threshold = visits │  betNo    │   ...    │ closes    │        │ (actualVisits)
  │ seedLiquidity()    │  by public│          │           │        │ processClaims()
  └──────────┬─────────┘     │     └──────────┴───────────┘        └──────┬──────┘
        oracle                public                                   oracle

Markets are identified by an 8-digit calendar key encoded as a uint256; the market for 27 June 2026 has id 20260627. This makes market identity human-legible, collision-free across days, and trivially sortable — properties we value more than opacity.

Let a market admit two outcomes, YES and NO. Bettors deposit the settlement token into the contract, contributing to a per-side pool. Write Y and N for the total YES and NO pools respectively, and T = Y + N for the total volume locked. There is no automated market maker and no scoring rule: stakes are simply pooled.

Odds. The implied probability of each outcome is its share of the total pool, reported in basis points (denominator 10 000):

Odds are endogenous: every bet moves the price against itself. A large YES stake raises p_YES, shrinking the marginal payout for subsequent YES bettors. This is the parimutuel discipline.

Resolution. Let v be the actual visit count and k the market threshold. The winning side is

Note the strict inequality: the YES side wins only if today’s visits exceed the threshold. Ties resolve NO. The threshold k is set, at market creation, to the visit count of the most recent completed local day — so each market asks the deceptively simple question, “more birds than yesterday?”

Payout. Let a winning bettor hold stake s on the winning side, whose pool totals W. The contract pays

That is, each winner reclaims their stake plus a proportional share of the entire losing pool. Summing (3) over the winning side recovers exactly T: the mechanism is fee-free and conservative — no token is created or destroyed at settlement, only redistributed.^a

Degenerate market refund. If a market resolves with an empty winning pool (W = 0) — for example, every bettor took the losing side — payout (3) is undefined. The contract handles this edge case explicitly: each participant is instead refunded their own stake on the opposing side, so no funds are stranded in escrow.^b

Truth enters the system through a single privileged oracle address. A daily automation executes the following idempotent steps, each guarded by a bearer secret:

1. create-next — opens tomorrow’s market with k equal to the last completed day’s count.
2. seedLiquidity — the protocol treasury stakes a capped amount on both sides, so the market has non-trivial opening odds before the first human arrives. Seeded liquidity is an ordinary stake and is itself subject to (3).
3. resolve-previous — submits the realized v for yesterday and fixes the winning side.
4. process-claims — paginates through the bettor set and pushes winnings on-chain, so users need take no action to be paid.

We acknowledge that a single-key oracle is a trust assumption of the first order. We make it deliberately. The cost of a decentralized oracle committee vastly exceeds the value at risk in a market about birds, and the failure mode of the centralized design — the operator lies about the count — is bounded, publicly auditable against the same feeder feed every user can watch, and, frankly, beneath the dignity of all parties involved.

Settlement uses MYKCLAWD, a standard ERC-20 on Base (0xE3C5FCfBfea42D5CE2492FD82c239B5503f17ba3). The contract is asset-agnostic — it accepts any ERC-20 as its bettingToken — but the production deployment hardcodes MYKCLAWD. There is no protocol fee, no rake, and no token emission tied to wagering; the only token flows are the parimutuel transfers of §4.

Bettors approve the contract once and place positions via betYes / betNo. Where the wallet supports it, approval and the bet are bundled into a single atomic EIP-5792 batch; otherwise the protocol falls back to two transactions. The economic security of a given market is therefore exactly the sum of stakes locked in it — nothing more, nothing less.

Truthful belief revelation. Under parimutuel payoff (3), a risk-neutral bettor maximizes expected return by staking on the side they believe undervalued relative to the current pool ratio. As volume grows, prices converge toward the crowd’s aggregate probability estimate — the standard information-aggregation result, here applied to chickadees.

The seed-supply attack. The settlement value v is a physical quantity that any actor can influence at low cost: one may simply add more birdseed, thereby attracting more visits and forcing YES. We consider three reasons this does not break the market.

• Bounded and noisy effect. The mapping from seed to visits is saturating and weather-dominated; a manipulator cannot guarantee crossing an arbitrary threshold.
• Symmetric availability. The attack increases visits but cannot decrease them, so it is a one-sided lever in a two-sided market; informed counterparties price it in.
• Welfare-positive. The dominant-strategy equilibrium of the seed-supply game is “more birds get fed.” We classify this not as an attack but as the protocol working as intended.

The settlement contract inherits standard hardening: reentrancy guards on every state-mutating external entry point, checks-effects-interactions ordering on claims, and SafeERC20 for all token movement. Market creation, resolution, and claim-processing are access-controlled to the oracle; treasury actions to the owner. Positions are marked claimed before transfer to prevent double-withdrawal.

The dominant residual risk is the oracle itself (§5). A secondary, purely operational hazard is observed in production: when the feeder is connected to a non-solar external power source, the upstream cloud API reports it OFFLINE and refuses to start the live stream, even while the device remains reachable and continues to upload visit data. This degrades the spectator experience but does not threaten settlement, since the counter entity is independent of the stream. The mitigation is to return the feeder to solar power, whereupon it recovers within seconds.

Visit counts are produced by consumer computer-vision hardware and are not exact; double-counts, missed visits, and the occasional confidently-misclassified squirrel are within the error budget. We make no claim that v is the ground-truth number of distinct birds — only that it is the number the feeder reports, which is the number the contract settles against. As the project itself states: the figures are almost certainly not 100% accurate, there will be bugs, and there will also be birds.

Several extensions are left to subsequent revisions: (i) multi-feeder markets with a composite index across geographically dispersed feeders; (ii) species-conditional markets (“will a cardinal appear before noon?”); (iii) a futures curve over weekly visit totals; and (iv) agent-mediated betting, already partially realized through the project’s Base MCP and Bankr integrations, in which autonomous agents read live odds and place positions on a user’s behalf.

We have shown that a real bird feeder can serve as a robust, daily-resolving oracle for an onchain parimutuel prediction market, and that the resulting system is simple enough to audit, cheap enough to operate, and absurd enough to enjoy. The birds, we reiterate, are unaffected. We commend the design to the community and to the feeder.