Research note · Market theory

Two ways to beat a market

Where excess return comes from when there is no ticker — and what the 67 markets we just screened have in common.

27 September 2026 Andrew Gree — founder & CEO, HyperC Research note

This is a research note about market structure, not a result. The 67 markets it discusses are proposed candidates, not catalogue entries; the only performance figures cited are from a labelled synthetic benchmark and do not establish live-market performance. Nothing here is a promise of profit. What is measured, and what is stated as unknown, lives on Research & evidence.

Two admissible answers

Every market on our new candidate list asks the same question: where does excess return come from when there is no ticker, no order book and no hedge fund? Flower clocks in Aalsmeer, GovDeals lots, B-Stock phone auctions, Faire catalogues, car-hauling load boards, component brokers, consignment counters — 67 of them, screened this week against nine criteria.

Finance theory allows only two answers. Either you produce the same outcome at a lower cost than the marginal participant — your trucks, cold room, licence, capital or exit channel are cheaper, so a deal that is negative for them is positive for you. Or you are right where other participants are systematically wrong — their estimates, their rules of thumb or their attention leave value on the table. Everything else is beta, luck or a fee.

Both answers are old. What is new is that in these markets the second one has a recorded, testable form: the history of deals that someone looked at and declined. That history is the raw material P34 is built to learn from, and this post is about what it contains.

The anatomy: transaction-rich, opportunity-sparse

All 67 markets share one shape. The menu is enormous — 18,930 distinct products on the Royal FloraHolland clocks, 15,000–25,000 live lots on GovDeals on any day, a billion part numbers on Sourcengine, 100,000 brands on Faire — and the subset that is profitable after every cost is small and scattered. Deal flow is not the scarce resource. Reliable screening is.

The decision-relevant value of an action a for an operator o at size q is not its gross spread but what is left after the operator's own costs and risks:

V(a, o, q) = G(a, o, q) − C_execution − C_operations − C_funding − E[L] − C_exit − ρ(a, o, q)

Read the terms and the two kinds of edge fall out of the formula. The cost terms — execution, operations, funding, exit — belong to the operator, not the market. A backhaul load is a loss for a carrier heading the wrong way and a profit for the one already going there; a pallet of restaurant equipment 40 miles away is worth more to a buyer with a truck and a showroom than to one who must hire both. Two operators can rationally assign different values to the same lot, and neither is wrong. That is comparative advantage, and it is where "be more efficient" lives.

The gross value G and the expected loss E[L] are estimates, and estimates are where other people are wrong. The winning bid on a mixed pallet is the most optimistic guess of what is inside; the closeout line nobody ordered is the line nobody priced. That is where "exploit their mistakes" lives.

The size term q is what makes the action space combinatorial. Buying 100 units and 20,000 units of the same SKU are different economic objects because of inventory ageing and liquidation; in a thin order book, $5,000 and $500,000 are different trades (Kyle, 1985). Thousands of nominal deals become millions of deal-size-timing choices, and a model that predicts unit economics but treats quantity as an afterthought will be profitable in a spreadsheet and destructive at scale.

The inefficiencies, in finance-theory terms

Seven textbook inefficiencies account for every reason a deal gets rejected in our catalogue taxonomy; four of them sit on the cost side and three on the estimate side.

Be cheaper to operate — your cost of the same outcome is lower	Be right where others are wrong — their estimate, rule or attention failed
Scale — ticket too small for the incumbent's overhead	Attention — nobody evaluated it: menu larger than the desk
Operational fit — bad for them, good for you: the backhaul	Prediction — somebody evaluated it and got it wrong
Capacity — no truck, no crew, no slot, no balance sheet	Complexity — too many interacting variables to price
Time — gone before the incumbent could act on it	Only this side leaves a recorded history of passes

P34 learns both from the same two tables: the money tape prices the costs; the reject frontier prices the mistakes. The left side is comparative advantage and needs no prediction to exploit; the right side is other people's errors and needs a model — but only the right side writes down what it passed on, which is why it is learnable.

Inefficiency (theory)	What it says	Reject reason	Where it shows up in our list
Costly information — Grossman & Stiglitz, 1980	Prices cannot fully reflect information that costs money to gather; someone must be paid to score 25,000 lots a day	Attention, Complexity	GovDeals (15,000–25,000 live lots), flower clocks (18,930 products), Faire (100,000 brands), component brokers (1B+ part numbers)
Limits to arbitrage — Shleifer & Vishny, 1997	Closing a mispricing needs capital, time and the capacity to hold; in physical markets it also needs trucks, cold rooms and licences, so mispricings persist	Operational fit, Capacity, Scale	Car-hauling boards, produce terminals, pallet recycling, restaurant-equipment auctions, hotel FF&E
Lemons and the winner's curse — Akerlof, 1970; Capen, Clapp & Campbell, 1971	With hidden quality the winning bid is the most optimistic estimate; uncorrected buyers overpay and sellers dump	Prediction	Unclaimed freight, insurance salvage, used-phone lots, ITAD lots, lab equipment
Rational inattention — Simon, 1955; Sims, 2003	A desk samples a fraction of the menu and never tests the rest; the untested majority is the reject set	Attention	Faire and cross-border catalogues, TikTok Shop product tests, consignment intake
Selection-biased histories — Heckman, 1979; Swaminathan & Joachims, 2015	Outcomes exist only for accepted deals, so a model trained on winners over-enters when the menu widens	Prediction (the estimate, not the opportunity, was the defect)	Every market on the list; sharpest in private-label launches, pawn books and credit-like decisions
Perishability (newsvendor) — Arrow, Harris & Marschak, 1951	Unsold stock is worth zero after the window; sizing under demand uncertainty is the whole decision	Time, Capacity	Flowers, fish, produce, hay, empty-leg flights, short-dated reagents
Market impact and strategy capacity — Kyle, 1985; Almgren & Chriss, 2001	Size moves the price and changes the remaining menu; returns are capacity-constrained	Scale, Capacity	Components in a shortage cycle, container trading, bulk wine, regional hay spreads

Seven families of markets, and what each one lives on

The 67 candidates fall into seven families, and each family monetises a different mix of the two edges. Access is open in most of them — that is deliberate. A market you cannot register for is not a market you can learn.

Family	Markets on the list	The inefficiency it lives on	Cost-side edge (be cheaper)	Estimate-side edge (be right)
Perishable auctions and spot wholesale	Royal FloraHolland clocks, Pefa fish clocks, Hunts Point produce, green-coffee spot lists, nursery stock, hay	Perishability and attention: the clock runs 100,000 transactions a day, nobody can hold, and the lot you passed at 07:10 is compost by Thursday	Cold chain, same-day dispatch, a customer base that turns stock in days	Which lot, at which price, on which day — grade-by-grade against your own realised sell-through
Surplus, liquidation and salvage	GovDeals, GSA and Public Surplus, restaurant-equipment auctions, lab and IT lots, B-Stock phone lots, unclaimed freight, insurance salvage, appliance truckloads, architectural salvage, trade-show teardowns, film props	Lemons and the winner's curse: sellers dump as-is, photos stand in for inspection, hard removal deadlines punish the buyer without a truck	Pickup radius, refurbishment bench, a resale channel that clears in weeks	Recovery from a manifest and a photo, corrected for the curse; refusing the lot with the title problem
Catalogue and reject-list markets	Faire and Ankorstore, cross-border marketplaces, private-label launches, TikTok Shop product tests, component brokerage, consignment intake, offline media remnant, IPv4 leases	Rational inattention: the menu is 100,000 brands or a billion part numbers and a buyer instantiates a fraction of it	Landed-cost, duty and returns handling; a store or channel that sells through	Scoring the whole catalogue including the lines nobody ordered — the reject set is the menu
Logistics capacity	Car-hauling load boards, inland barge quotes, chassis and trailer pools, container trading and one-way leasing, pallet and tote recycling, recovered-material bales, empty-leg charter	Operator-dependent value: the same load is a loss for one carrier and a backhaul for another; the menu expires in hours and one accept changes the neighbours	Truck or barge position, trailer type, depot network, lane relationships — the edge is almost entirely operational	Which combination of loads to accept under routing constraints; damage, detention and cancellation risk
Industrial feedstocks and remnants	Printing-paper stocklots, mill-end textiles and film remnants, biomass chip and pellet auctions, DDGS and ethanol co-products, used cooking oil routes, spent catalyst	Matching cost: a roll, a lot or a stream is worth more to the one small job it fits than to the channel that rejected it, and moisture, spec or assay is only known on delivery	Haul distance, conversion or sampling capacity, a buyer pipeline of small jobs	Spec-match and quality estimation; refusing the long-haul lot whose freight eats the remnant value
Licensed and gated markets	Pawn lending, state-legal cannabis wholesale, bulk and fine wine, aviation parts, medical equipment, ag-input dealers, carbon credits, PGM reclaim	Limits to arbitrage by licence: the moat keeps efficient capital out, so incumbents run on policy tables written years ago	Holding the licence, the bonded premises or the accreditation is the edge; P34 does not confer it	Replacing the incumbent's rule table with expected economics — the same move as our micro-lending pilot, and the same gated launch
Venture-class bets	Private-label launches, TikTok Shop tests, small-business acquisitions, Reg CF and syndicate deals	Selection-biased histories: every operator remembers the launches that worked, and the observed outcomes come only from the deals they took	Sourcing, QC and ad operations that lower the cost of a failed test	Learning not to make the false-positive bets that sink the portfolio; TikTok tests are the one family where a bet costs hundreds and resolves in days

The families differ in which side dominates. Logistics capacity is almost all cost-side: a load board is efficient for everyone in aggregate and inefficient for you in particular. Surplus auctions are almost all estimate-side: the truck matters, but the money is made or lost in the bid. Perishable auctions demand both at once, which is why they sit at the top of our fit ranking and why almost nobody runs them with a model.

How P34 monetises each side

P34 takes two tables and a description, and each of the two edges is learned from a different table.

Operational alpha comes from the money tape. The Sales table is not list prices; it is what was actually realised — per-unit price, fee, holding cost, returns, the write-off when the leftovers were never sold. Because the replay reconstructs profit from that tape, the model learns your V(a,o,q), not the market's. A carrier's tape teaches it that a Dallas–Phoenix load at $1.10 a mile is worth taking on a Tuesday and not on a Friday; a flower wholesaler's tape teaches it which grades its customers will still pay for on day three. No comparable-sales database contains that, because it is not a fact about the flower. It is a fact about the operator.

Judgment alpha comes from the reject frontier. The Menus table carries every option that was available at each decision moment, taken and untaken, with a flag for what the prior policy chose. That flag is not an outcome label; it is a behaviour-policy signal. It tells the model where the old desk's threshold sat, which is exactly where the mistakes cluster: the offer nobody evaluated (attention), the offer evaluated with the wrong estimate (prediction), the offer too tangled to price (complexity). P34 constructs several plausible interpretations of the market, discards the dominated ones, and treats no-trade as an action that earns its place in the portfolio.

Refusing is the larger half. In a low-margin market a handful of false positives — lots predicted profitable that realise a loss after hidden downside — dominate the year. That is why the model is scored on portfolio profit and false-positive exposure, not on row-level error, and why the common output on a real menu is a short list of buys and a long list of passes. In our synthetic slower-market-waves benchmark, the same menu that bled a tuned gradient-boosting regressor to −$62.6k over 20 holdout dates left P34 at +$25.8k, almost entirely by declining what the regressor took. Synthetic, controlled, not a promise — but the mechanism is the point: the edge on the estimate side is mostly the losses you did not take.

The two sides also compound. A cheaper operator can afford to bid on deals the market prices at zero, so its history contains outcomes the market never observed; that history sharpens its estimates; better estimates let it size up where its cost advantage is real. Comparative advantage buys the data that judgment needs.

Which inefficiencies decay, and which persist

Not all seven reject reasons age the same way, and the order in which they erode decides where a member should stand.

Attention erodes first. Any agent can now read 25,000 GovDeals lots or a 100,000-brand catalogue before breakfast. As agents proliferate, the pure attention reject — nobody looked — shrinks toward zero, and with it the easiest profit. Graded coins show the end state: a free archive of millions of realised prices, open registration, no consumer, and a round-trip fee that eats whatever selection edge was left. Clean telemetry plus open access plus nothing physical is a market that has already been solved.

Prediction erodes as data pools. The estimate-side reject persists only while histories stay private. A member's own tape is private by construction, which is why the model is fitted per operator rather than once for the world; but the winner's-curse premium in a salvage lot will narrow as more bidders correct for it. Expect this edge to be real for years and shrinking.

Operational fit, capacity, scale and time persist. These are limits to arbitrage in the Shleifer–Vishny sense, and a better model does not move them. A load is still wrong for a truck heading the other way; a pallet still has to be collected by Thursday; a bank still cannot economically write a $283 loan against a drill. As long as the marginal participant has to own a truck, a cold room or a licence, the reject set it leaves behind is structural.

The strategic implication is uncomfortable for a pure software company and comfortable for an operator: the durable edge is model plus operations, tuned to a part of the market the operator already serves. That is why the fit ranking rewards markets where anyone can register but few can execute — and why a member who claims a market and instruments it early holds an advantage that later entrants cannot download.

Where the thesis stops

The screen found as many boundary cases as fits, and they are worth naming because each one is a way the computable-market thesis can be falsified for a given market.

Too transparent. Australian wool is objectively tested before sale and every passed-in lot is published — perfect declined-option telemetry and almost no mispricing left to learn. Recovered-material bales price off a monthly index. The edge there is freight, not selection.
One irreducible shock. A used mining rig is worth whatever the hashprice says this morning; a bulk-wine lot is worth whatever a multi-year oversupply decides. When one factor the operator does not control dominates the outcome, a conditional model adds little over a one-line rule.
Manipulated telemetry. Game keys that get revoked, grey-market lots with counterfeits mixed in, "tested" hardware that was not. If strategic participants write the labels, refusal is most of the value and the label itself cannot be trusted; only component brokerage has an institutional fix (AS6081 testing, ERAI reporting).
Feedback too slow. Timber sales scale out over months, small-business acquisitions over years, Reg CF investments over a decade. The paper's condition is that feedback must arrive faster than the regime changes; these markets fail it for a single operator and survive only with pooled histories.
Closed access. Hotel bed-bank inventory is a perfect menu — hotel × date × room type — behind contracts that forbid resale of net rates. Structure without access is a case study, not a market.
Licensed perimeters. Pawn, cannabis, alcohol, medical devices, securities and environmental commodities clear the structural test and sit behind a licence or a regulator. They belong on the enterprise track as gated launches, with counsel first and no profit-share pricing, exactly as our micro-lending pilot runs today.

Two caveats stand over everything above. The evidence for P34 is a synthetic benchmark plus one production deployment in Amazon wholesale, company-reported; nothing here is a forecast of returns in any of the 67 markets. And no-free-lunch results still hold: the thesis is that under adequate support, attribution and feedback a decision system can approach the best policy measurable from the operator's own records — not that compute can recover information the records never contained.

What we are doing about it

We publish evidence, not promises. Every market that moves from blueprint to experiment goes through the same sequence: reconcile the operator's accounting, replay history with date-aware controls, stress-test against margin compression and missing labels, shadow-score live menus with no execution, then a capped pilot. Rejection rates, predicted-versus-realised economics, wins, losses and failure cases go on the record as they happen.

The 67 markets in this screen are now on the catalogue's waiting list at blueprint state, ranked by structural fit and discounted by perimeter. The top of the ranking is where anyone can register and few can execute: government surplus, the flower clocks, Faire-class catalogues, restaurant-equipment auctions, cross-border marketplaces, used-phone lots, produce terminals. The bottom is where structure is perfect and something else is missing — access, a consumer, or feedback fast enough to learn from.

If you operate in one of these markets, or in the 136th we have not listed, the useful first step is not the model. It is the snapshot: start recording the whole menu you see each day, including everything you pass on. That record is the asset. The model is how you read it.

Don't believe us. Compute it.

Sources

HyperC, the computable markets catalogue and markets.json — the 135 existing entries, the four criteria and the seven reject reasons
HyperC, P34 API documentation — Menus, Sales and market_type contract; data floors; the candidate waiting list
HyperC, Computable Markets working paper and technical report — value decomposition, boundary conditions, synthetic benchmark (−$62.6k vs +$25.8k over 20 holdout dates)
Royal FloraHolland, 2025 results and FloralDaily on clock turnover — 18,930 products, >100,000 transactions/day
GovDeals, by the numbers; BidProwl, GovDeals FAQ — 15,000–25,000 live lots
Sourcengine, API integrations — 1B+ part numbers; ERAI counterfeit statistics via Oxebridge
Faire, how Faire works — 100,000+ brands
Central Dispatch, carrier network; CCJ, 2026 car-shipping rates
FirstCash, FY2025 10-K — average pawn loan $312
AWEX, wool auctions; Hotel Tech Report, bed banks
Theory: Grossman & Stiglitz (1980); Shleifer & Vishny (1997); Akerlof (1970); Capen, Clapp & Campbell (1971); Simon (1955); Sims (2003); Heckman (1979); Swaminathan & Joachims (2015); Arrow, Harris & Marschak (1951); Kyle (1985); Almgren & Chriss (2001) — linked in the table above

The full scored list (67 markets, nine criteria, editable weights, plus a reconciliation of 25 further proposals) is the companion workbook P34_candidate_markets.xlsx and the machine-readable candidate_markets.json.

Free

Follow the model as it develops

Research notes as we write them, open problems as we hit them, model and agent-tool releases and member findings. Free, and you can leave whenever you like.

By subscribing you agree to the Privacy Notice. Unsubscribe anytime.

Research & evidence

The working paper, benchmark methodology, notebook results and the limitations we state rather than hide.

See the evidence →

More notes

Research, open problems, releases and what we are learning about computable markets.

All notes →

Research note — 27 September 2026. Proposed candidates are not catalogue entries; synthetic evidence is labeled as synthetic and does not establish live-market performance.