A 2021 betting bot, the full on-chain round history, and a reproducible post-mortem of a +19% paper edge that nets out to roughly minus the fee.
Every number and figure below is regenerated by python -m research.pipeline; the dataset invariants are asserted by tests, and CI re-runs both on every push. The repo keeps its 2021 name — the study below explains why "arbitrage" earned its scare quotes.
PancakeSwap Prediction is a parimutuel bet on five-minute BNB/USD moves: back Bull or Bear, winners split the losers' pool minus a 3% fee. In 2021 I noticed the two sides' payouts often diverged wildly and built a bot on a simple hypothesis: if a five-minute crypto move is a coin flip, betting the thin side whenever its implied edge cleared +25% should be free money — the average round quoted a +19% edge under that assumption. Across 143,856 settled rounds, the assumption fails in the most instructive way possible: the more a side pays, the less often it wins (43.3% for underdogs overall, down to 23% in the extreme tail), tracking the payout-implied breakeven rate almost one-for-one. What survives is a small reverse favorite–longshot bias — underdogs at final odds returned +5.5% per bet above the +25% threshold — but it lived in thin pools, faded as the market deepened, could not absorb even 1 BNB of size, and was never observable at bet time anyway. The one durable lesson is about transaction costs: a five-minute directional edge far too small to trade on an exchange becomes a several-percent bet once a binary payoff amplifies it (~500× in the data), which is why informed money bothers to price this pool at all — and, once it does, why the arbitrage layered on top cannot survive. I ran the strategy live in 2021; this repo contains the bot, the data, and the pipeline that measures why the paper edge was a mirage.
pip install -r requirements.txt
python -m research.pipeline # rebuilds results/stats.json, results/edge_bins.md, figures/
pytest # strategy math, dataset invariants, pipeline smoke testThe pipeline is deterministic, runs in about half a minute on a laptop, and writes every statistic cited here to results/stats.json in machine-readable form. To watch the bot reason about live rounds without risking anything:
python -m bot.run # dry-run by default: logs would-be bets, signs nothingPancakeSwap Prediction (the PancakePredictionV2 contract on BNB Chain) runs staggered five-minute rounds around a Chainlink BNB/USD feed:
start ──────── betting open ──────── lock ──────── price runs ──────── close
| bet Bull or Bear (~5 min) | oracle price recorded (~5 min) |
└── round n's betting window overlaps round n−1's live window ──────────┘
Bull wins if the close price exceeds the lock price, Bear if it is lower; on an exact tie the house keeps the whole pool (387 rounds, 0.27% of the sample). A winning 1 BNB stake pays the parimutuel multiplier
m = (1 − fee) · totalAmount / winnerPool
stake included. The treasury fee, measured from the data rather than assumed, is 3.00% in every settled round with a winning side. Median pool: 23.0 BNB per round; median round length: 309 seconds.
If every round were a fair coin flip, the expected net return of one unit on side S would be EV = 0.5·m − 1, which is positive whenever S holds less than 48.5% of the pool. Pools are lopsided often: the average round's underdog quoted a naive edge of +19%, and 54% of rounds quoted at least +10%.
This is exactly the signal the 2021 bot traded: watch the pools, and in the final ten seconds before lock fire a bet whenever the naive EV cleared +25%. The rest of this document is about why that number did not survive contact with the data — the more the pool "pays" you to take a side, the less likely that side is to win.
data/rounds.csv.gz holds all 144,000 rounds served by the contract's public rounds() view between epoch 100 (2021-08-26) and epoch 144,099 (2023-02-06). Filters: 119 voided rounds (oracleCalled = FALSE) and 25 rounds with an empty side are excluded, leaving 143,856. Amounts were exported in scientific notation (~6 significant digits — exact enough for payout ratios, not for wei-level accounting); prices are Chainlink integers ×1e8; timestamps UTC. The sample includes 40 service gaps longer than an hour (chain halts and contract pauses; the longest, 58 hours, in November 2021). Column dictionary and caveats: data/README.md.
Two artifacts of the 2021 live run survive in assets/legacy/: the bot's ex-post EV distribution and the first win-rate regression, kept as provenance for §6.
Sort each round's underdog into bins of implied edge and compare three numbers per bin: the win rate its final odds require to break even, the win rate it achieved, and the realized return per bet (auto-generated at results/edge_bins.md):
| Implied edge (naive EV) | Rounds | Mean payout | Breakeven win rate | Actual win rate | Realized return/bet | t |
|---|---|---|---|---|---|---|
| <0 | 17,803 | 1.97× | 50.8% | 49.5% | −2.46% | −3.3 |
| 0–0.1 | 48,479 | 2.09× | 47.8% | 47.0% | −1.78% | −3.7 |
| 0.1–0.25 | 41,362 | 2.33× | 43.0% | 43.0% | +0.05% | +0.1 |
| 0.25–0.5 | 24,138 | 2.69× | 37.2% | 37.9% | +2.02% | +2.4 |
| 0.5–0.75 | 7,077 | 3.21× | 31.3% | 33.0% | +5.49% | +3.1 |
| 0.75–1 | 2,591 | 3.71× | 27.0% | 30.4% | +12.44% | +3.7 |
| 1–1.5 | 1,672 | 4.38× | 22.9% | 27.0% | +17.72% | +3.7 |
| ≥1.5 | 734 | 6.53× | 16.6% | 23.2% | +69.12% | +3.6 |
| all | 143,856 | 2.38× | — | 43.3% | +0.50% | — |
Realized returns are computed at final odds — an infeasible benchmark: they condition on the closing pool, which no bettor can observe before committing.
The 50% column the 2021 hypothesis predicted does not exist. Win rates fall from 49.5% to 23.2% in near-lockstep with the breakeven rate — the calibration figure above puts the fitted slope at 0.83 (SE 0.02), and a logit of underdog wins on implied edge rejects the random walk at any conventional level (coefficient −0.84, z = −19.5; heteroskedasticity-robust standard errors here and throughout). The pool ratio at lock is a price, and even in a five-minute crypto casino, prices aggregate information.
A slope below one means the crowd slightly under-prices extreme underdogs — the reverse of the racetrack favorite–longshot bias. At final odds, backing every underdog above the bot's +25% threshold returned +5.53% per bet (t = 6.6, n = 36,212); above +50%, +12.5% (t = 6.8). Two things stop this from being an investment thesis.
First, "final odds" is an infeasible benchmark — it conditions on the closing pool, which no bettor can observe before committing. Second, the premium did not survive the market's first year: +28% per bet in 2021Q3 fades to −1% by 2022Q2 as the market matured and the median pool behind those bets deepened from 21 toward 34 BNB, and it re-emerges (+6% in 2022Q4, +8% in the partial 2023Q1) only once the dying market had thinned back to 13 BNB pools. Whether arbitrageurs competed it away or deeper pools simply diluted it, the data cannot say — what they show is that the mispricing never persisted where real money could follow it.
The cross-section agrees, if noisily: within the τ ≥ 0.25 sample, returns fall from +11.2% per bet in the smallest-pool quartile (median 11 BNB) to +4.0% in the largest (43 BNB), non-monotonically in between (+6.2%, +0.7%).
An itemized account of the gap between the quoted edge and a live P&L, at the bot's own threshold (τ ≥ 0.25):
| Step | Edge per bet (0.1 BNB stake where size matters) |
|---|---|
| Implied by the 50/50 assumption at final odds (mean over selected bets — a proxy for the bet-time quote, ~10s of flow away) | +51% |
| … after the win-rate correction (win 35.6%, not 50%), at final odds | +5.5% |
| … after your own 0.1 BNB dilutes the thin pool you're betting into | +4.0% |
| … after gas (~0.001 BNB per bet + amortized claim at 2021's 5 gwei; estimate) | ≈ +3% |
| … after the last-seconds flow you cannot see (you commit ~10s before lock) | ≈ 0 to −3% |
| At 1 BNB of size, dilution alone | −4.6% |
The last row before size is the one the on-chain snapshot cannot decompose — final pools only — but the 2021 live run recorded it directly: the ex-post EV measured after lock sat systematically below the EV quoted at bet time (legacy figure, mass well below the ex-ante threshold). Informed money arrives last, and in a parimutuel market the last money sets your odds. The live strategy netted out around minus the fee: an efficient-markets result in the venue where you would least expect one.
The 2021 design, preserved in spirit: poll the contract, and in the last ten seconds before lock compute each side's EV from the live pools; fire betBull/betBear when the threshold clears; re-read the pools ~15 seconds after lock to log the ex-post EV (the piece that later falsified the hypothesis); batch claim() calls to amortize gas. The rewrite modernizes the engineering while keeping that logic:
bot/strategy.py— the decision math as pure functions, including the exact fee treatment (0.5·(1−fee)·T/S − 1; the 2021 code'sT/S/2 − 1.03treated the fee as additive, overstating EV precisely on the imbalanced rounds it selected) and own-stake dilution. Fully unit-tested, no network imports.bot/client.py— a thin web3 v7 client: nonce from the pending pool, gas price from the node (BNB Chain's floor is 0.1 gwei since 2025, not 2021's hardcoded 5), claims discovered viagetUserRoundsinstead of one call per historical epoch.bot/run.py— sleeps until each round's betting window instead of busy-polling the RPC, retries with capped exponential backoff, and defaults toDRY_RUN=true: it signs and sends nothing until you pass--live.bot/config.py+.env.example— credentials live in an untracked.env, never in source. Use a burner wallet if you go live.
Do not fund this bot — the central finding of this repository is that it does not make money. It works; the strategy doesn't. Run it dry to watch it lose hypothetically, or run python -m research.pipeline to see why it must. (The contract itself is still live: epoch ~496,000 as of July 2026.)
Why could anyone predict BNB over five minutes? The resolution is that nobody does. The final pool share predicts the round's direction strongly (logit coefficient 3.5, z = 63), and it subsumes momentum: last-round return predicts direction on its own (z = 11.5), but conditional on the pool share its sign flips and shrinks (z = −2.1). Bettors chase momentum (corr(share, r₋₁) = 0.22), some of the flow is informed, and the informed part arrives late — the adverse-selection logic of parimutuel models with late informed bettors (Ottaviani & Sørensen, 2009), which accommodates either sign of the resulting bias. The result is a market where odds track outcomes at slope 0.83: not perfectly efficient — extreme longshots stayed cheap in the thin early months, the reverse of the racetrack favorite–longshot bias catalogued by Thaler & Ziemba (1988) and revisited by Snowberg & Wolfers (2010) — but efficient enough that fees, impact, and timing ate the residual for any real bettor. A five-minute casino on a chain full of bots is still a market.
Section 8 leaves a supply-side question: if the pool aggregates information, who puts it there, and why here rather than on the exchange where BNB actually trades? The answer turns on transaction costs, and it is the most general thing this project has to say.
A trader who can call the direction of a five-minute BNB move even slightly better than chance holds a signal that is worthless on a spot or perpetual exchange. The five-minute return has a standard deviation of 0.21% and a mean absolute size of 0.14% (base_rates in results/stats.json), so a taker acting on, say, a 55%-accurate directional call expects (2·0.55 − 1) × 0.14% ≈ 1.4 basis points of gross P&L per trade — below the exchange's taker fee and spread. An edge that small is collected only by the market maker who earns the spread; the taker cannot afford to act on it, so to them the price is an unpredictable martingale.
The prediction market pays for the sign of the move and throws the magnitude away. That same 55% call on a roughly balanced pool returns 0.55 × 1.94 − 1 ≈ +6.7% per bet — the identical signal, amplified about 500-fold, with gas costing a cent. The amplification is not an artifact of the example: at zero fee the binary bet returns exactly 2p − 1 while the exchange returns (2p − 1) × E|move|, so the factor is 1 / E|move| ≈ 700× no matter how large or small the edge. A binary payoff is leverage on the sign of the move, and it strips out precisely the ~0.14% of magnitude that transaction costs live inside.
So the venue exists, and its odds are informative, because it converts sub-transaction-cost directional edges into bets worth placing — and the informed money that shows up to collect that amplified edge is the same late, adverse-selecting flow that prices the pool to slope 0.83 and erases the imbalance "arbitrage" this project began with. The two markets are efficient for opposite structural reasons: the exchange because a five-minute edge cannot clear costs there — Jensen's (1978) transaction-cost definition, where a signal is only an edge if you can trade it net of costs — and the prediction market because, once its binary payoff makes that same edge cheap enough to trade, enough people trade it that the odds price it in. A microscopic inefficiency invisible in the order book becomes a visible, and then arbitraged-away, price in the casino next door.
- The dataset snapshots final pools only. The late-flow mechanism in §6 is inferred from the 2021 live logs and the decay pattern, not decomposed bet-by-bet on-chain. Scraping
BetBull/BetBearevent logs would timestamp the flow inside each window and identify when the smart money arrives — the natural next study. - All backtests ignore execution latency; returns are per-bet means, not a funded equity curve. The headline t-stats treat rounds as independent, but the conclusion does not lean on that: clustering by calendar day leaves the τ ≥ 0.25 t-stat at 5.9 across 526 days (
backtest_final_oddsinresults/stats.json). - Single asset (BNB/USD), single venue, single fee regime, Aug 2021 – Feb 2023. The residual-premium estimates are tail-sensitive: the ≥1.5 bin holds 734 rounds with 6.5× mean payouts, so a handful of outcomes moves it by percentage points. None of these caveats can resurrect the 50/50 assumption, which fails at z = −19.5 across the full sample.
- The 5-minute return autocorrelation (+0.08, t = 31) partly reflects the oracle's update mechanics; I make no market-microstructure claims about BNB itself.
├── data/ # rounds.csv.gz (144k rounds) + column dictionary
├── research/ # loaders → metrics → figures → pipeline (one command)
├── results/ # stats.json + edge_bins.md, regenerated by the pipeline
├── figures/ # the four figures above, regenerated by the pipeline
├── bot/ # strategy (pure math) / client (web3) / run (CLI) / config
├── tests/ # strategy math, dataset invariants, dry-run safety, smoke
└── assets/legacy/ # the 2021 live-run figures, kept as provenance
The original 2021 main.py and keys.py are preserved in git history. MIT license. If you use the dataset, cite via CITATION.cff.
Built live on BNB Chain in 2021; revisited as a study in 2026.