Skeptical Evaluation: Three Short-Horizon Prediction-Market Bot Strategies

Polymarket 5m / 15m Crypto Direction Markets

Reviewer posture: every strategy is guilty of not working until proven otherwise.

---

Structural Facts That Constrain Everything

Before evaluating any strategy, the execution environment imposes hard constraints that no amount of clever signal generation can overcome.

Market mechanics. Polymarket operates a hybrid-decentralized CLOB (Central Limit Order Book): matching is offchain, settlement is onchain (Polygon), and all orders are limit-order primitives — "market orders" are just marketable limits (FAK/FOK).docs.polymarket.com Outcome tokens redeem to $\$1$ or $\$0$ at resolution, so all PnL is bounded binary payoff math plus execution frictions.docs.polymarket.com

Displayed price ≠ your fill. The displayed market price is the midpoint of best bid and best ask — unless the spread exceeds $\$0.10$, in which case the UI shows last traded price instead. Buying happens at the ask; selling at the bid; depth determines your real fill. Any backtest using displayed price as fill price is suspect.docs.polymarket.com

Fee structure. Polymarket uses a probability-based dynamic taker fee. For crypto markets (including 5m and 15m), the documented formula is:

$$\text{fee per share} = C \times \text{feeRate} \times p \times (1-p)$$

where $C$ is the number of shares, $p$ is the share price, and the crypto taker feeRate is $0.072$. The peak effective rate is 1.80% at $p = 0.50$. Makers pay zero fees and may receive rebates (20–25% of taker fees for crypto).docs.polymarket.com +1

The fee curve at operationally relevant price points:

Bottom line: At every price level, you need a positive calibration edge over the market price just to break even against taker fees. There is no free lunch at any price.

Fee regime changes matter. As of March 30, 2026 ("Fee Structure V2"), taker fees apply to ten market categories including all crypto timeframes (5m, 15m, 1H, 4H, daily, weekly). The fee rate and category assignments have changed multiple times in 2025–2026. Any backtest must be segmented by fee regime.docs.polymarket.com

Volume and competition. 5-minute crypto markets draw up to $60 million daily volume, with bots accounting for 55–62% of that volume.blockchain.news +1 You are competing against purpose-built latency-optimized bots from day one.

Historical data limitations. The official price history endpoint provides aggregated bars by interval/fidelity; real microstructure analysis requires your own recorded tick-by-tick book and trade stream via websocket.docs.polymarket.com +1

Execution state is not instant. Trade lifecycle includes MATCHED, MINED, CONFIRMED, RETRYING, and FAILED. If your bot measures fills only at signal time and ignores status changes, your live PnL accounting will be wrong. Rate-limit throttling can also queue requests and silently add latency.docs.polymarket.com

---

STRATEGY 1: Copy-Trading / Shadowing a Top Wallet

Core Thesis

Someone else has already solved the hard problem — building a model, getting fast data, calibrating probabilities. You find them by screening wallet PnL histories on the public blockchain, then mirror their trades with proportional sizing.

Why It Might Work in Prediction Markets

On-chain transparency is the structural reason this idea even exists. Everything is on the public Polygon blockchain, meaning every trade by every wallet is observable. Several platforms have built infrastructure around this.quicknode.com If a target wallet genuinely has persistent edge from a superior model or faster data feed, and that edge exceeds your execution degradation, copying can theoretically extract a fraction of their alpha.

Main Failure Modes

1. Latency decay (likely fatal). Total latency from on-chain event detection to your confirmed trade: typically 4–14 seconds under normal Polygon network conditions, and potentially 10–30 seconds for a retail builder without a dedicated node. In a 5-minute contract (300 seconds), 10–30 seconds is 3.3–10% of the contract's entire life. By the time you detect the trade, the order book has already absorbed the information. You enter at a strictly worse price than the target.quantvps.com

At 30-second latency into a 5-minute contract, empirical modeling suggests you are buying a contract that has already moved against you by approximately +1.1 percentage points in adverse price.

2. Survivorship bias in target selection (severe). You select wallets by screening past PnL. But past PnL is a mixture of skill and luck. With thousands of wallets active, some will have impressive track records purely by chance. The strongest-looking wallets are the most likely to revert. Only ~12–13% of users are net profitable overall.medium.com

Academic evidence from social trading platforms confirms that lagging, follower-type trades underperform leading trades, and that peer-performance chasing increases activity while reducing returns and increasing volatility.sciencedirect.com

3. Maker vs. taker asymmetry. If the target places limit orders (maker), they pay zero fees and may earn rebates. You, reacting to their trade, are almost certainly a taker. You pay 0.65–1.80% per trade that the target does not. Your net PnL is structurally worse than theirs even with zero latency.

4. Crowding. Multiple copy-trading platforms now exist. More bots chasing the same wallets collectively move the book against each other.financemagnates.com

5. Unobservable intent. A "top wallet" trade may be part of a wider portfolio hedge, maker inventory adjustment, or cross-market strategy. You see the entry but not the context. You cannot observe exit signals or model recalibrations.

6. Target adapts or goes dormant. The wallet you've spent weeks validating may change strategies, move to different markets, or stop trading entirely.

How Latency Affects It

Extreme sensitivity. Even an additional 50ms of latency can make profitable trades unviable in high-frequency contexts. Documented retail bot builders have reported slippage of 6+ cents per trade — enough to erase most edges.reddit.com

For 15-minute contracts, latency matters less proportionally (30s / 900s = 3.3% vs 30s / 300s = 10%), which is why if you build this at all, it should only be tested on 15-minute contracts.

How Fees and Slippage Affect It

Devastating. You pay full taker fees on every copy trade. The target may pay zero (maker). On a $p = 0.50$ entry, your fee is 1.80% of position value. On top of this, slippage from delayed execution adds another 0.5–2%. Combined, you need the target's raw alpha to exceed ~3–4% per trade just to break even after your costs — an absurdly high bar for 5-minute markets.

Robust for Automation?

Technically simple (WebSocket monitoring → order submission). Operationally fragile. The bot works, but the edge doesn't. Verdict: weak as standalone strategy. Better as a feature in a broader model.

Recommended Mode

Shadow mode / research only. Shadow the target for 500+ trades. Simulate your fills with realistic latency and fee models. If the shadow P&L is negative, kill the idea before risking capital.

Data Required

• Full on-chain Polygon transaction history for candidate wallets (minimum 6 months)
• Orderbook snapshots at the instant of the target's trade (to simulate your fill)
• Exact timestamp of when your system could first observe the wallet action
• Your own latency logs from day one of shadow mode
• Per-contract win rate, stake, timing, and market conditions

Metrics to Track

Minimum Sample Size

• ≥1,000 observed leader trades across ≥3 distinct market regimes (trending up, trending down, ranging)
• ≥300 executable delayed-copy simulations per leader
• across 60+ trading days and multiple volatility regimes
• with positive lower 95% confidence bound on net expectancy after clustering by coin-day

What Would Make Me Reject It Entirely

• Target's alpha disappears when simulating even 3 seconds of delay
• Target's P&L is concentrated in <10 big wins (luck, not skill)
• Multiple other copy bots already trail this wallet
• You cannot achieve maker fills on any leg
• Copied trades are net negative after taker fees and realistic slippage
• Edge exists only at midpoint/last-trade prices, not executable quotes

---

STRATEGY 2: Coin Spread / Relative-Value Across Markets

Core Thesis

BTC-Up and ETH-Up contracts on the same 5-minute window are correlated because crypto moves together. When one coin's market is priced "too cheap" relative to its historical relationship with the other, buy the cheap side and sell the rich side. Profit when the spread mean-reverts, regardless of which direction crypto actually moves.

Why It Might Work in Prediction Markets

Prediction markets quote separate binary books for related underlyings. Because books are shallow and update asynchronously, temporary cross-market dislocations can occur even when the underlying spot relationships are tight.sciencedirect.com Price discovery in crypto genuinely leads in BTC and follows in altcoins on short horizons. Market makers may not reprice all contracts simultaneously, creating brief windows of stale prices. Relative-value also reduces dependence on absolute directional forecasting.

Main Failure Modes

1. Not true arbitrage (critical conceptual flaw). You cannot construct a true pairs trade in binary prediction markets. In equities, you can long Stock A and short Stock B to be delta-neutral. In Polymarket, you can buy BTC-UP and buy ETH-DOWN, but neither creates a genuine hedge. "BTC goes up" and "ETH goes up" are not fungible events — they're correlated, not identical. Even with $\rho = 0.8$, there is a 20% residual that can go either way. The "relative value" framing creates false comfort.

2. Double fees (likely fatal). You need two legs: buy one contract, sell/buy another. Both are taker trades. At $p \approx 0.50$ (maximum uncertainty zone where spreads are widest), each leg costs ~1.80%. Round-trip fee: ~3.60%. You need a spread divergence exceeding 3.60% just to break even. If the typical divergence is 2–5%, fees eat most or all of the edge.

3. Legging risk. You must fill both legs near-simultaneously. If leg 1 fills and leg 2 fails (order cancelled, book moved, gas spike), you are left with unhedged directional exposure — the opposite of what you wanted.

4. Correlated blowup risk is severe and quantifiable. With 5 positions at $\rho = 0.85$ (typical crypto correlation):

$$P(\text{all 5 lose}) \approx 5.9\%$$

compared to:

$$P(\text{all 5 lose, independent}) = (0.15)^5 = 0.008\%$$

That is approximately 780× higher catastrophic-loss probability due to correlation. When a BTC macro event triggers a sudden reversal, all positions resolve against you simultaneously.

5. Lead-lag is a sub-second phenomenon. The genuine price-discovery lead of BTC over altcoins operates on millisecond-to-low-second timescales in spot markets. By the time a Polymarket contract price reflects the BTC spot move and you can detect, parse, and respond, the altcoin contract has already repriced.

6. Thin books on the minor leg. BTC is liquid; SOL or DOGE is not. The minor leg will slip more, and the slip is always against you.

How Latency Affects It

Extreme, and uniquely dangerous. Simultaneous execution of both legs is the defining requirement. A 2-second gap between legs means the second leg fills at a different price than modeled. For a solo builder without colocated infrastructure and sub-second order execution, 5-minute contracts are likely not feasible. For 15-minute contracts, the lead-lag window is somewhat wider, but still very short.

How Fees and Slippage Affect It

Likely fatal for taker execution. Two taker legs at 1.0–1.8% each gives 2.0–3.6% round-trip. If your median spread capture is 2%, you're structurally losing. The only escape is placing limit orders (maker, zero fees) on at least one leg — but then you sacrifice execution speed and risk partial fills.

Robust for Automation?

The most complex of the three. Requires real-time cross-market spread calculation, simultaneous order management, inventory tracking across positions, correlation estimation, regime detection, and failure-mode handling for partial fills. High complexity = high bug risk. A solo builder will spend weeks on infrastructure before generating a single signal.

However: if the execution and risk controls are first-class, the mathematical framework is objective and testable — making it the strategy with the strongest structural foundation in theory.

Recommended Mode

Research first, then shadow, then extremely cautious live. You need to prove the spread exists, persists after double fees, and can realistically be captured before building anything.

Important variant (confidence: high across models): Rather than building a true two-leg spread, consider a BTC-signal → altcoin-entry single-leg approach:

"BTC is X% into a 15-minute window and has moved Y standard deviations from open. Buy ETH-direction contract if ETH contract price implies a lower probability than BTC spot move suggests."

This eliminates the legging risk and double-fee problem while retaining the actual signal.datawallet.com

Data Required

• Synchronized L2 orderbook snapshots for all coin pairs (tick-level)
• Trade prints and timestamps for both legs
• External underlying spot/reference prices (Pyth/Chainlink)
• Historical cross-coin correlation matrices by market regime
• Per-market fee rates fetched dynamically (never hardcoded)docs.polymarket.com
• Fill simulator incorporating partial fills, queue decay, cancel latency

Metrics to Track

Minimum Sample Size

• ≥500–1,000 completed paired trades (both legs executed)
• ≥100 live or latency-injected paper trades
• across 3+ regimes: quiet, trend, violent reversal
• at least 30–60 active trading days
• positive lower 95% confidence bound on net pair expectancy

What Would Make Me Reject It Entirely

• Double fees exceed median spread capture
• Correlation between target coins drops below 0.5 on 5-min horizons
• Legging failures occur >30% of the time
• Spread half-life exceeds 80% of contract duration
• Net PnL strongly tracks outright BTC/ETH direction (secretly just directional)
• Edge exists only on midpoints, not on executable depth
• Order book depth on altcoin contracts is consistently under $200 top-of-book

---

STRATEGY 3: Buying High-Probability Contracts (0.80–0.90)

Core Thesis

Systematically buy contracts priced at 0.80–0.90. Win most of the time, collecting the 10–20 cent spread to resolution at $\$1.00$. Accept an asymmetric payoff (small frequent wins, rare large losses) and profit from systematic miscalibration — the market prices these outcomes slightly below their true probability.

Why It Might Work in Prediction Markets

The favorite-longshot bias is a well-documented empirical regularity in betting markets: longshots are overbet whereas favorites are underbet.journals.uchicago.edu In crypto direction markets, this would manifest as the "Down" side being overpriced when BTC is trending upward (people love betting on crashes), leaving the "Up" side at 0.80–0.90 slightly underpriced relative to true probability. Prediction-market prices map naturally to probabilities, and academic work suggests prices can be biased, particularly near the extremes.users.nber.org

Additionally, the fee structure creates a structural advantage: fees are lowest at the extremes. At $p = 0.90$, the effective fee rate is just 0.72%, compared to 3.60% at $p = 0.50$. This means the fee hurdle is lowest precisely where this strategy operates.

Main Failure Modes

1. No calibration edge — the null hypothesis (fundamental). If the market price of 0.85 reflects a true probability of exactly 85%, your expected value is:

$$\text{EV} = 0.85 \times 0.14082 - 0.15 \times 0.85918 = 0.1197 - 0.1289 = -0.0092$$

That's −0.92 cents per share. Without a genuine calibration edge over the market, this strategy is strictly negative EV after fees. The fee is the house edge. A contract at $0.85 means the market believes the probability is 85%. If the market is efficient, there is no edge. The appearance of edge in backtests comes entirely from the fact that 85% of these contracts do indeed resolve at $1.00 — which is exactly what you paid for.

2. Convex loss structure (the steamroller). At $p = 0.85$, you lose 6.1× what you win on each trade. At $p = 0.90$, it's 9.7×. This means a single loss wipes out 6–10 wins. A small miscalibration (true prob = 83% instead of 85%) turns a "winning" strategy into a slow bleed. The payoff structure punishes drawdown severely.

3. Correlated multi-position losses (the kill shot). If you simultaneously hold BTC-Up at 0.85, ETH-Up at 0.87, and SOL-Up at 0.82, and a market-wide crash occurs, all positions lose at once. Under independent assumptions, the probability of all 5 positions losing is 0.008%. Under realistic crypto correlation ($\rho \approx 0.7$–$0.85$), it rises to approximately 5–6% per window — not a tail event, a semi-regular operational reality that transforms a "safe" strategy into one with periodic 400+ USDC drawdowns on modest sizing.

4. Last-second oracle/resolution risk. Approximately 15–20% of 5-minute BTC windows involve material price movements in the final 10 seconds.medium.com A contract at $0.88 in the 4th minute can move to $0.15 in the 5th minute due to a sudden price reversal. This risk cannot be backtested accurately without genuine last-second tick data.

5. Adverse selection from the sell side. Who is selling you these "high probability" contracts? In an efficient market: participants who have information that the probability is lower than 0.85. Market-makers who maintain these prices may be selling to you precisely because they have tighter latency-sensitive pricing signals you do not.

6. Overtrading and exposure stacking. A high win rate creates behavioral and algorithmic overconfidence. You keep adding positions because "it always works." This stacks exposure and accelerates the correlated blowup when it comes.

How Latency Affects It

Low-to-moderate, compared to the other two. Unlike copy-trading or spread trading, you're not racing anyone for fills. You can place limit orders (maker, zero fees) early in the contract's life and wait for a fill. The exception: if you try to enter late when the outcome appears nearly certain, the book will be thin and you'll pay large spreads.

Critical caveat: If conditions deteriorate in the 4th minute, you may want to exit. Thin order books mean your exit price is far below your entry. Latency to monitor and respond determines whether you exit at $0.65 or $0.40.

How Fees and Slippage Affect It

Significant but manageable — IF you have edge. At $p = 0.85$, the breakeven true probability is 85.9%. You need to be better calibrated than the market by +0.92pp. At $p = 0.90$, breakeven is 90.6% (need +0.65pp edge). These are not impossible hurdles, but they're not trivial either.

Critical insight: If you can place maker orders instead of taker orders, your fee is zero (and you may earn rebates). This would eliminate the fee hurdle entirely and make any positive calibration edge exploitable. The trade-off: maker orders may not fill, especially if the market moves away from your price.

Beyond fees: Documented slippage on retail bots has been 6+ cents per trade in thin markets.reddit.com On a strategy that earns only 14¢ per winning share, that's up to 43% of gross profit consumed before considering fees.

Robust for Automation?

The simplest of the three. Logic: monitor market → check if price is in target range → check correlation exposure cap → place order → track resolution. No cross-market coordination, no target wallet dependency, no simultaneous multi-leg execution.

However, the automation logic must enforce:

• Correlated exposure limits (no more than 1–2 positions per BTC-correlated group)
• Time-in-window restrictions (no entry after minute 3 of a 5m contract, minute 9 of a 15m contract)
• Exit-on-deterioration logic (require active monitoring, not just hold-to-resolution)

Recommended Mode

Shadow mode first, then micro-live. This is the most backtest-fakeable strategy (see comparison below), which makes shadow validation essential.

• Shadow 500+ qualifying contracts over 4–8 weeks
• Log every "would-have-entered" trade with timestamp, price, and resolution
• Compute observed win rate vs. market-implied probability
• Only go live if observed edge exceeds breakeven by ≥2× the standard error

Data Required

• Per-contract price time series at 1-second resolution, aligned to window open/close timestamps
• Entry-side best ask, not midpoint (this is critical and most builders miss it)
• Full order book depth around entry time
• Pyth/Chainlink price feed snapshots synchronized to contract prices (to verify oracle drift)
• Last-30-seconds price data per window — this is the key data most builders lack
• Volatility regime label for each window (trending / ranging / reversal)
• Historical resolution outcomes with timestamps
• Dynamic fee schedule per marketdocs.polymarket.com

Metrics to Track

Minimum Sample Size

Because this strategy hides risk in rare losers, you need enough losses to estimate the tail:

To detect a 2pp calibration edge (e.g., true prob = 87% when market says 85%) at 95% confidence:

Do not trust any result with fewer than 50–100 observed losses. A strategy that claims 90% win rate but has only had 10 losses is not validated — it's untested.

At 5-min markets running ~288 contracts/day per coin (24h × 12/hr), for 3 coins that's ~864 contracts/day. But you won't enter every contract — only when price is 0.80–0.90. Realistically 10–30% qualify. Expect 4–8 weeks minimum for adequate shadow data.

What Would Make Me Reject It Entirely

• Observed win rate at $p = 0.85$ is ≤ 85.9% over 1,000+ trades (no edge after fees)
• Loss correlation across coins exceeds 0.5 (correlated blowup is frequent)
• Max drawdown exceeds 3× the daily variance expectation
• Most profit disappears when using ask instead of midpoint
• Kelly fraction is so small (<0.5%) that realistic position sizes cannot cover minimum order sizes
• No time-of-day, volatility regime, or coin filter improves calibration
• PnL is dominated by a few lucky regimes

Kelly Criterion Context

If you genuinely believe the probability is 87% while the market says 85%, at the effective win/loss payoff:

$$f_{\text{Kelly}} = \frac{p_{\text{true}} - p_{\text{be}}}{1 - p_{\text{be}}}$$

At $p = 0.85$, breakeven = 0.8592, true $p$ = 0.87:

$$f_{\text{Kelly}} \approx 0.077 \quad (7.7\%)$$

Use quarter-Kelly (divide by 4): ~1.9% of bankroll per trade. For a $1,000 account, that's $19 per trade. Full Kelly is suicidal with estimated probabilities.

---

Comparative Analysis

Edge Durability

Copy-Trading: Edges decay as copy-bots proliferate. Wallet alpha is typically regime-specific and exhausted within weeks of discovery. Estimated half-life: 2–4 weeks post-selection.

Coin Spread / RV: If structural (BTC leads, altcoins lag), potentially more durable since it monetizes microstructure inconsistency rather than pure forecasting. But correlation arbitrage windows shrink as more sophisticated actors enter. Estimated half-life: 1–3 months, but likely never reachable by retail in its pure two-leg form.

High-Prob: The structural miscalibration (if it exists) is the most stable, since it depends on persistent market-maker behavior and retail psychology rather than another trader's secret edge. However, it is also the shallowest source of edge. Estimated half-life: 3–6 months, conditional on the calibration anomaly existing at all.

Sensitivity to Delay

Copy-Trading and Coin Spread are both fatally sensitive to delay for 5-minute contracts. A 15–30 second lag converts any theoretical edge into a guaranteed negative. High-Prob is the only strategy where a 15–30 second monitoring lag does not immediately destroy the edge, though it does affect exit capability.

Risk of Correlated Blowups

Most dangerous for Strategy 3 if position limits are not enforced. A naive bot buying BTC-Up, ETH-Up, and SOL-Up simultaneously in the same window faces a joint loss probability of ~5–6% per window due to crypto correlation. Over 100 windows per day, this is not a tail event — it's a recurring operational reality.

Most dangerous for Strategy 2 if it's secretly just directional exposure wearing a hedge costume.

---

Direct Answers to Your Final Comparison Questions

Safest One to Test First

Strategy 3 (High-Prob Contracts), on 15-minute contracts only, one coin at a time, with a maker-only entry rule.

Confidence: high (4 out of 6 models agree).

Reasoning: simplest to implement, clearest rejection criteria, lowest infrastructure cost, produces interpretable and attributable PnL. You can generate data within 1–2 weeks and have a statistically meaningful answer within 6–8 weeks. Even if it fails, you've learned what calibration edges look like in this market. A failed shadow test tells you something actionable.

Dissenting view (2 models): Strategy 2 (Coin Spread) in shadow mode is the safest because it can reduce outright market beta and forces you to confront real execution from day one. The counterargument is that its implementation complexity creates more ways to introduce bugs and false signals before you learn anything.

Highest Upside

Strategy 2 (Coin Spread / Relative Value), specifically the single-leg BTC-signal → altcoin-entry variant.

Confidence: high (5 out of 6 models agree).

If — and this is a large "if" — a persistent cross-market information asymmetry exists that survives execution costs and can be captured without perfect simultaneity, the edge is market-neutral, repeatable, and potentially the most scalable. The single-leg variant (use BTC price action as a signal, trade the lagging altcoin's direction contract) eliminates the double-fee problem and legging risk.

Easiest to Fake-Good in Backtests But Fail Live

Strategy 3 (High-Prob), by a meaningful margin. Strategy 1 (Copy-Trading) is a close second.

Confidence: high (broad agreement, though models disagree on which is #1).

Why Strategy 3 fakes well: A backtest of "buy all contracts priced 0.80–0.90, hold to resolution" will show ~85%+ win rate and positive P&L on historical data. This looks compelling. But it will have all of the following flaws invisibly:

• No slippage (you got the midpoint, not the ask)
• No fee modeling (or simplified flat fee instead of the dynamic formula)
• No last-second reversal data (missing from most historical datasets)
• No order book depth check (assuming you could always fill)
• No correlation between positions (treating each trade as independent)
• Look-ahead bias in regime selection

Why Strategy 1 fakes well: You screen 10,000 wallets, find the top 10 by PnL, backtest copying them, and get spectacular results — because you selected for the outcome. Live, the wallet's edge may have been luck, or it may have decayed, or the book may have thinned from other copiers.

Rule: Never trust a backtest of any of these strategies unless it was built on tick-level data with explicit order book depth checks, fee simulation using the exact dynamic formula, and last-second reversal modeling.

Most Realistic Staged Rollout Plan

Skeptical Evaluation: Three Short-Horizon Prediction-Market Bot Strategies

Polymarket 5m / 15m Crypto Direction Markets

Reviewer posture: every strategy is guilty of not working until proven otherwise.

---

Verified Market Parameters (April 2026)

Before evaluating anything, the fee regime and microstructure constrain what is theoretically possible.

Polymarket operates a hybrid-decentralized CLOB (Central Limit Order Book): matching is offchain, settlement is onchain (Polygon), and all orders are signed limit-order primitives — "market orders" are just marketable limit behavior (FAK/FOK).docs.polymarket.com +1

Displayed price is not your fill. The UI shows the midpoint of best bid and best ask, unless the spread exceeds $\$0.10$, in which case it shows last traded price. Buying happens at the ask; selling at the bid; depth determines your real fill. Any backtest using displayed price as fill price is invalid.docs.polymarket.com

Outcome tokens redeem to $\$1$ or $\$0$ at resolution — all PnL is bounded binary payoff math plus execution frictions.docs.polymarket.com

Fee structure: Polymarket uses a probability-based dynamic taker fee. The documented formula is:

$$\text{fee} = C \times \text{feeRate} \times p \times (1-p)$$

where $C$ is share count, $p$ is share price, and feeRate for crypto markets is 0.072 (post Fee Structure V2, March 30, 2026). Peak effective taker fee is 1.80% at $p = 0.50$. Makers pay zero fees and may receive rebates (20–25% of collected taker fees for crypto).docs.polymarket.com +2

Fee rates should never be hardcoded — official docs explicitly state they can vary and must be fetched dynamically.docs.polymarket.com

5-minute crypto markets launched February 12, 2026 with taker fees enabled, following the same fee curve as 15-minute markets. Starting March 6, 2026, taker fees extend to all crypto markets (1H, 4H, daily, weekly).docs.polymarket.com

The computed fee table at relevant price points (post-V2):

Bottom line: At every price level, you need a positive calibration edge over the market price just to break even against taker fees. There is no free lunch at any price.

Additional structural facts:

• 5-minute markets draw up to $60M daily volume; bots account for 55–62% of volume.blockchain.news +1
• Polymarket introduced dynamic fees specifically to curb latency arbitrage.financemagnates.com
• Gas fees on Polygon are typically under $\$0.01$ per transaction.
• Documented retail bot slippage can reach 6+ cents per trade in thin markets.reddit.com
• GTD order expiration has a one-minute security threshold — you cannot rely on auto-expiry for last-second protection.docs.polymarket.com
• Trade lifecycle includes MATCHED, MINED, CONFIRMED, RETRYING, and FAILED states.docs.polymarket.com
• Historical price endpoints are aggregated by interval/fidelity; for 5m/15m markets, you need your own recorded tick-by-tick book/trade stream.docs.polymarket.com

---

STRATEGY 1: Copy-Trading / Shadowing a Top Wallet

Core Thesis

Someone else has solved the hard problem — building a model, getting fast data, calibrating probabilities. You find them by screening wallet PnL histories on the public blockchain, then mirror their trades with proportional sizing.

Why It Might Work in Prediction Markets

On-chain transparency is the structural reason this idea even exists. Every trade by every wallet is observable on Polygon. If a target wallet genuinely has persistent edge from a superior model or faster data feed, and that edge exceeds your execution degradation, copying can theoretically extract a fraction of their alpha.

Public market data, onchain data, and subgraphs make wallet behavior study possible — rarer in traditional venues.docs.polymarket.com

Main Failure Modes

1. Latency Decay (likely fatal). Total latency from on-chain event detection to your confirmed trade: typically 4–14 seconds under normal Polygon conditions; minimum realistic round-trip for retail infrastructure is 5–8 seconds, with 10–30 seconds being common.quantvps.com In a 5-minute contract (300 seconds), 10–30 seconds is 3–10% of the contract's entire life. By the time you detect the trade, the order book has already absorbed the information. You enter at a strictly worse price.

2. Survivorship Bias in Target Selection (severe). You select wallets by screening past PnL. But past PnL is a mixture of skill and luck. With thousands of active wallets, some will have impressive records purely by chance. The strongest-looking wallets are the most likely to revert. Academic evidence from social trading platforms finds that peer-performance chasing increases activity while reducing returns and increasing volatility.sciencedirect.com

3. Maker vs. Taker Asymmetry. If the target places limit orders (maker), they pay zero fees and earn rebates. You, reacting to their trade, are almost certainly a taker. You pay 0.65–1.80% per trade that the target does not. Your net PnL is structurally worse than theirs even with zero latency.

4. Unobservable Intent. A "top wallet" trade may be part of a wider portfolio hedge, maker inventory adjustment, or cross-market book management — not a standalone directional bet. You cannot see their full context.

5. Crowding. Multiple copy-trading platforms now exist.quicknode.com More bots chase the same edge → spreads tighten → latency becomes decisive → opportunity shrinks or disappears. When several copiers pile into the same wallet, they collectively move the book against each other.

6. Target Adapts or Goes Dormant. The wallet you've validated may change strategies, move to different markets, or stop trading entirely.

7. Observation Delay. Matching is offchain and settlement is onchain. If you infer leader behavior from onchain activity or indexed subgraphs, you are observing after the economically useful moment.docs.polymarket.com

How Latency Affects It

Extreme sensitivity. Even 50ms additional latency can make profitable trades unviable for high-frequency strategies. At 4–14 seconds of copy delay, you are not competing — you are picking up scraps. If the target's edge is 2% and the price moves 1.5% in the seconds after their trade, you capture 0.5% gross, minus your taker fee. That's negative.newyorkcityservers.com +1

For 15-minute contracts, latency matters less proportionally (30s / 900s = 3.3% vs 30s / 300s = 10%), so if you build this, it should only be tested on 15-minute contracts.

How Fees and Slippage Affect It

Devastating. You pay full taker fees on every copy trade. The target may pay zero (maker). Combined with slippage from delayed execution (documented at 6+ cents per trade for retail bots in thin markets), you need the target's raw alpha to exceed ~3–4% per trade just to break even.reddit.com

Robust for Automation?

No, not as blind auto-copying. Technically simple (WebSocket monitoring → order submission). Operationally fragile — the bot works, but the edge doesn't. At best, wallet activity is a research feature or a shadow signal, not a standalone execution strategy.

Recommended Mode

Shadow mode / research only. Shadow the target for 500+ trades. Simulate fills with realistic latency and fee models. If shadow P&L is negative, kill the idea before risking capital.

Data Required

• Full on-chain Polygon transaction history for candidate wallets (minimum 6 months)
• Polymarket CLOB order book snapshots at 1-second granularity for entry price reconstruction
• Underlying spot prices (Pyth/Chainlink) synchronized to block timestamps
• Your own detection latency logs from day 1 of shadow mode
• Per-contract win rate, stake, timing, and market conditions
• Exact timestamp of when your system could first observe the wallet action

Metrics to Track

Minimum Sample Size

At least 1,000 observed leader trades across 60+ trading days and multiple volatility regimes (trending up, trending down, ranging). At least 300 executable delayed-copy simulations per leader. With positive lower 95% confidence bound on net expectancy after clustering by coin-day.

To detect a 3 pp edge (true $p = 0.53$ vs null $p = 0.50$) at 80% power and 5% significance: ~1,716 trades. At 20 live copy trades per day, this is ~86 trading days (~4 months).

What Would Make Me Reject It Entirely

• Target's alpha disappears when simulating even 3 seconds of delay
• Target's P&L is concentrated in <10 big wins (luck, not skill)
• Multiple other copy bots already trail this wallet
• You cannot achieve maker fills on any leg
• Target trades fewer than 20 times per week
• Edge exists only at midpoint or last-trade prices, not executable quotes
• Wallet activity cannot be observed early enough without privileged data
• Net expectancy ≤ 0 after conservative costs

---

STRATEGY 2: Coin Spread / Relative-Value Across Markets

Core Thesis

BTC-Up and ETH-Up contracts on the same 5-minute window are correlated because crypto moves together. When one coin's market is priced "too cheap" relative to its historical relationship with the other, buy the cheap side and sell the rich side. Profit when the spread mean-reverts, regardless of overall crypto direction.

Why It Might Work in Prediction Markets

Prediction markets quote separate binary books for related underlyings. Because books are shallow and update asynchronously, temporary cross-market dislocations can occur even when spot relationships are tight. Academic work on cryptocurrencies documents intraday cross-predictability and lead-lag effects across coins.sciencedirect.com

Additionally, different liquidity providers for different coins may not reprice all contracts simultaneously, creating brief windows of stale prices. Relative-value reduces dependence on absolute directional forecasting.

Main Failure Modes

1. Double Fees (likely fatal). You need two legs: buy one contract, sell (or buy the opposite side of) another. Both are taker trades. At $p \approx 0.50$, each leg costs ~1.80%. Round-trip fee: ~3.60%. You need a spread divergence exceeding 3.60% just to break even. If the typical divergence is 2–5%, fees eat most or all of the edge.

2. Legging Risk (severe). You must fill both legs near-simultaneously. If leg 1 fills and leg 2 fails (order cancelled, book moved, gas spike), you are left with unhedged directional exposure — the opposite of what you wanted.

3. Not True Arbitrage. "BTC goes up" and "ETH goes up" are correlated, not identical events. Even with $\rho = 0.8$, there is a ~20% residual. In a 5-minute window, idiosyncratic coin-specific news can easily decorrelate them. Subtle rule/oracle/time differences between contracts add further basis risk.

4. Thin Books on the Minor Leg. BTC contracts are relatively liquid; SOL or DOGE is not. The minor leg will slip more, and the slip is always against you.

5. Correlation Instability. Crypto correlations are regime-dependent. They spike to ~0.95 during panic selloffs (destroying your hedge) and drop to ~0.3 during idiosyncratic moves (making the spread meaningless).

6. Lead-Lag Is a Sub-Second Phenomenon. The genuine price-discovery lead of BTC over altcoins operates on millisecond-to-low-second timescales. By the time a Polymarket contract reflects the BTC spot move and you can detect, parse, and respond, the altcoin market has already repriced.

7. Hidden Beta Stacking. You think you're market-neutral, but you're actually long crypto risk cluster. During a flash crash, all positions resolve against you simultaneously.

Correlated Blowup Analysis

This risk deserves quantitative emphasis. With 5 positions at $\rho = 0.85$ (typical crypto correlation):

$$P(\text{all 5 lose}) \approx 5.9\%$$

compared to:

$$P(\text{all 5 lose, independent}) = (0.15)^5 = 0.008\%$$

That is approximately 780× higher catastrophic-loss probability due to correlation. When BTC experiences a macro event — exchange hack, regulatory news, liquidation cascade — all positions resolve against you simultaneously. This is not diversification; it is concentrated exposure in a diversification costume.

How Latency Affects It

Extreme. Simultaneous execution is the defining requirement. A 2-second gap between legs means the second leg fills at a different price than modeled. For a solo builder without colocated infrastructure and sub-second order execution, this strategy is likely not feasible for 5-minute contracts. For 15-minute contracts, the lead-lag window is somewhat wider but still very short.

How Fees and Slippage Affect It

Likely fatal for taker execution. Two taker legs at 1.0–1.8% each gives 2.0–3.6% round-trip. If your median spread capture is 2%, you're structurally losing. The only escape is placing limit orders (maker, zero fees) on at least one leg — but then you sacrifice execution speed and risk partial fills.

Robust for Automation?

Highest complexity of the three. Requires real-time cross-market spread calculation, simultaneous order management, inventory tracking, correlation estimation, regime detection, and failure-mode handling for partial fills. A solo builder will spend weeks on infrastructure before generating a single signal.

Confidence note: Models disagree on whether this strategy is the safest to test first. Several models (GPT-5.3, GPT-5.4, Gemma, Mistral) favor Strategy 2 in shadow mode as safest because it reduces directional beta. Other models (Claude Opus, Claude Sonnet, Grok) favor Strategy 3 as safest because it's simpler. The disagreement is ~50/50 across models. My synthesis: Strategy 2 is theoretically safer if you can solve execution, but Strategy 3 is practically safer because it's simpler to validate and reject.

Recommended Mode

Research first, then shadow. You need to prove the spread exists, persists after double fees, and can realistically be captured before building anything.

Important reframing from GPT-5.4: If you build this, consider building it as a directional signal extraction tool (not as a spread trade): "BTC has moved Y standard deviations from open → buy ETH-direction contract if ETH contract price implies a lower probability than BTC spot move suggests." This eliminates the two-leg execution problem while retaining the signal.docs.polymarket.com

Data Required

• Synchronized L2 order books for both legs at sub-second resolution
• Trade prints + timestamps for both legs
• External underlying spot/reference prices (Pyth, Chainlink)
• Contract rule metadata (end time, oracle, resolution logic)
• Fill simulator with partial fills, queue decay, cancel latency
• Historical cross-coin correlation matrices by market regime
• Dynamic fee rates per market

Metrics to Track

Minimum Sample Size

At least 500–1,000 completed paired trades, across at least 2–3 volatility regimes. Partial-fill trades must be included (they're part of reality). At least 200 observations per coin pair. At least 100 live or latency-injected paper trades. At 10 complete spread trades per day: ~96 trading days minimum.

What Would Make Me Reject It Entirely

• Double fees exceed median spread capture
• Legging loss dominates gross convergence alpha
• Correlation between target coins drops below 0.5 on 5-min horizons
• Partial fills occur >30% of the time
• Spread half-life exceeds 80% of contract duration
• Results require constant retuning of pair relationships
• Residual directional beta is large (strategy is secretly just outright direction)
• Order book depth on altcoin contracts consistently under $\$200$ top-of-book
• Net edge vanishes after two-sided costs

---

STRATEGY 3: Buying High-Probability Contracts (0.80–0.90)

Core Thesis

Systematically buy contracts priced at 0.80–0.90. Win most of the time, collecting the 10–20 cent spread to resolution at $\$1.00$. Accept an asymmetric payoff (small frequent wins, rare large losses) and profit from any systematic underpricing of favorites.

Why It Might Work in Prediction Markets

The well-documented favorite-longshot bias describes the empirical regularity that long shots are overbet while favorites are underbet.nber.org +1 In crypto direction markets, this could manifest as the "Down" side being overpriced when BTC is trending upward, leaving the "Up" side slightly underpriced.

Prediction-market prices map naturally to probabilities, and academic work suggests prices often aggregate beliefs fairly well, but mispricing near the extremes can occur.users.nber.org

Additionally, the fee structure itself creates an asymmetry: fees are lowest at the extremes. At $p = 0.90$, the effective fee rate is just 0.72%, compared to 3.60% at $p = 0.50$. The fee hurdle is lowest precisely where this strategy operates.

Main Failure Modes

1. No Calibration Edge (the null hypothesis). If the market price of 0.85 reflects a true probability of exactly 85%, your expected value is:

$$\text{EV} = 0.85 \times 0.1408 - 0.15 \times 0.8592 = 0.1197 - 0.1289 = -0.0092$$

That's −0.92 cents per share. Without a genuine calibration edge over the market, this strategy is strictly negative EV after fees. The fee is the house edge.

2. Convex Loss Structure (the steamroller). At $p = 0.85$: you lose 6.1× what you win on each trade. At $p = 0.90$: 9.7×. One loss wipes out 6–10 wins. A small miscalibration (true prob = 83% instead of 85%) turns a "winning" strategy into a slow bleed.

3. Correlated Multi-Position Losses (the kill shot). If you simultaneously hold BTC-Up at 0.85, ETH-Up at 0.87, and SOL-Up at 0.82, and a market-wide crash occurs, all positions lose at once. Under independent assumptions, the probability of all 5 positions losing is ~0.008%. Under realistic crypto correlation ($\rho \approx 0.85$), it approximates ~5.9% — not a tail event but a semi-regular operational reality.

4. Last-Second Oracle Risk. ~15–20% of 5-minute windows involve significant price action in the final 10 seconds. A contract at 0.88 can move to 0.15 in seconds due to a sudden reversal. This risk cannot be backtested accurately without genuine last-second tick data.

5. Adverse Selection from the Sell Side. Who is selling you contracts at 0.85? In an efficient market: participants who have information that the probability is lower. Market makers maintaining these prices may be selling precisely because they have tighter latency-sensitive pricing signals.

6. Overtrading. A high win rate creates behavioral overconfidence. You keep adding positions because "it always works." This stacks exposure and accelerates the correlated blowup when it comes.

How Latency Affects It

Low-to-moderate sensitivity compared to the other two strategies. You're not racing anyone for fills. You can place limit orders (maker, zero fees) early in the contract's life and wait for a fill. The exception: if you try to enter late when the outcome appears nearly certain, the book will be thin and you'll pay large spreads. If conditions deteriorate in the 4th minute, you may want to exit, but thin books mean your exit price is far below entry.

How Fees and Slippage Affect It

Significant but manageable — IF you have edge. At $p = 0.85$, the breakeven true probability is 85.9%. You need to be better calibrated than the market by +0.92 pp. At $p = 0.90$, breakeven is 90.6% (need +0.65 pp edge). These are not impossible hurdles, but they're not trivial either.

Critical insight: If you can place maker orders instead of taker, your fee is zero (and you may earn rebates). This eliminates the fee hurdle entirely and makes any positive calibration edge exploitable. The trade-off: maker orders may not fill, especially if the market moves away from your price.

On a 5m or 15m horizon with tiny per-trade margins, slippage of even 1–2¢ on a 14¢ potential profit is 7–14% additional degradation.

Robust for Automation?

Simplest of the three. Logic: monitor market → check if price is in target range → check correlation exposure cap → place order → track resolution. No cross-market coordination, no target wallet dependency, no simultaneous multi-leg execution.

However, the automation logic must enforce strict rules on correlated exposure limits, time-in-window restrictions, and exit-on-deterioration logic.

Recommended Mode

Shadow mode first, then micro-live.

• Shadow 500+ qualifying contracts over 4–8 weeks
• Log every "would-have-entered" trade with timestamp, price, and resolution
• Compute observed win rate vs. market-implied probability
• Only go live if observed edge exceeds breakeven by ≥ 2× the standard error

Data Required

• Per-contract price time series at 1-second resolution, aligned to window open/close
• Full order book depth (bid/ask/size) at entry time — midpoint is not sufficient
• Underlying reference-price path at sub-second resolution (Pyth/Chainlink)
• Time-to-expiry field for all open positions
• Last-30-seconds price data per window (this is the key data most builders lack)
• Volatility regime label for each window
• Historical resolution outcomes with timestamps
• Dynamic fee schedule per marketdocs.polymarket.com

Metrics to Track

Minimum Sample Size

To detect a 1.5 pp edge (true $p = 0.874$ vs break-even $p = 0.859$) at 80% power: ~3,310 trades. At 20 trades per day: ~166 trading days (~5.5 months).

To detect a 2 pp edge at 95% confidence:

Do not trust any result with fewer than 50–100 observed losses. A strategy claiming 90% win rate with only 10 losses is not validated — it's untested. Because per-trade edge is small and the payoff distribution is highly skewed, a massive sample is required to observe the true frequency of catastrophic tail events.

What Would Make Me Reject It Entirely

• Observed win rate at $p = 0.85$ is ≤ 85.9% over 1,000+ trades (no edge after fees)
• Loss correlation across coins exceeds 0.5 (correlated blowup is likely)
• Max drawdown exceeds 3× the daily variance expectation
• Kelly fraction is so small (<0.5%) that realistic position sizes cannot cover minimum order sizes
• No time-of-day, volatility regime, or coin filter improves calibration
• Shadow P&L is negative after 200 trades even before fee simulation
• More than 2 instances of "late reversal" (contract at 0.85+ reverting below 0.50 in final 60 seconds) per 100 trades
• Most profit disappears when using ask price instead of midpoint
• Net calibration ≤ 0 in two independent out-of-sample periods

---

Comparative Analysis

Expected Edge Durability

Copy-Trading: Edges decay as copy-bots proliferate. Wallet alpha is typically regime-specific and exhausted within weeks of discovery. Estimated half-life: 2–4 weeks post-selection.

RV Spread: If structural relationships persist, the underlying mispricing mechanism is more durable than behavioral copying. But double fees remain a permanent drag, and any persistent sub-second lead-lag opportunity will be competed away by colocated actors. Estimated half-life: 1–3 months, but possibly never reachable by retail.

High-Prob: The structural miscalibration (if it exists) depends on persistent market-maker behavior and retail psychology rather than another trader's secret edge. The shallowest source of edge but potentially the most stable. Estimated half-life: 3–6 months, conditional on the calibration anomaly existing at all.

Implementation Complexity

• Copy-Trading: Requires Polygon node or high-frequency API, on-chain transaction parsing, CLOB order mapping, wallet watchlist. ~400–600 hours of engineering.
• RV Spread: Synchronized multi-market data feeds, implied fair value model, multi-leg execution, correlation regime detection. ~600–1,000 hours.
• High-Prob: CLOB price monitoring for one contract at a time, fee calculation, exposure limits. ~100–200 hours for a functional v1.

Sensitivity to Delay

Copy-Trading and RV Spread are both fatally sensitive for 5-minute contracts. A 15–30 second lag converts any theoretical edge into a guaranteed negative. High-Prob is the only strategy where a 15–30 second monitoring lag does not immediately destroy the edge.

Risk of Correlated Blowups

The RV Spread strategy has the worst catastrophic downside if it encourages holding 3–5 correlated positions simultaneously. With $\rho = 0.85$, $P(\text{all 5 lose}) \approx 5.9\%$ per window — not a tail event, a semi-regular operational reality.

Strategy 3 has similar correlated blowup risk if unsized — but it is the easiest to cap with strict position limits.

---

Proposals

Safest One to Test First

Strategy 3 (High-Prob Contracts), on 15-minute contracts only, one coin (BTC), with maker-only entry.

Model disagreement note: ~50% of models favor Strategy 2 (RV Spread) in shadow mode as "safest" because it reduces directional beta. The other ~50% favor Strategy 3 because it's simpler to implement, validate, and reject. My synthesis: Strategy 3 is the safer practical choice for a solo retail builder because it requires less infrastructure, produces interpretable/attributable P&L, and its failure modes are the most clearly identifiable. Strategy 2 is theoretically safer if you can solve execution, but that "if" is doing enormous load-bearing work.

Why Strategy 3:

• Simplest to implement correctly
• Clearest rejection criteria
• Lowest infrastructure cost
• Generates data within 1–2 weeks
• Produces a statistically meaningful answer within 6–8 weeks
• Even if it fails, you learn what calibration edges look like in this market
• A failed shadow test tells you something actionable

Highest Upside

Strategy 2 (Coin Spread / Relative Value) — if and only if you solve execution and maintain true neutrality. Structural relationships may be more durable than pure directional or copy alpha.

Important reframing: Build it as a BTC-signal-to-ETH-entry single-leg directional trade, not a two-leg spread. This eliminates the "correlation spread" delusion while retaining the actual signal.

Easiest to Fake-Good in Backtests But Fail Live

Strategy 3 (High-Prob Contracts), by a large margin. Runner-up: Strategy 1 (Copy-Trading).

Why Strategy 3 fakes well:

• Backtests show ~85%+ win rate and positive P&L on historical data
• No slippage modeled (you assume midpoint fills)
• No fee modeling or simplified flat fee
• No last-second reversal data (missing from most datasets)
• No order book depth check (assumed unlimited fills)
• No correlation between positions (each trade treated as independent)
• Look-ahead bias in regime selection

Why Strategy 1 fakes well:

• You screen 10,000 wallets, find the top 10 by PnL, backtest copying them — spectacular results because you selected for the outcome
• Historical leader fills are not your fills
• Cherry-picked wallets and periods

Rule: Never trust a backtest of Strategy 3 unless it was built on tick-level data with explicit order book depth checks, the exact dynamic fee formula, and last-second reversal modeling. Never trust a backtest of Strategy 1 unless the wallet selection was done prospectively (out-of-sample).

Most Realistic Staged Rollout Plan

Strategy 1 is not recommended for any live phase. The latency penalty in 5-minute markets makes it structurally nonviable for a solo retail builder without colocated infrastructure. Keep it as a research feature only — wallet activity as one input to a broader model.

---

Design Rules for a Production Bot

These rules assume you are running Strategy 3 (high-probability contracts) on Polymarket 15m crypto direction markets. They are designed to prevent the specific failure modes identified above. Treat violations as system defects requiring immediate remediation, not judgment calls to override manually.

---

Rule 1: When to Enter

Enter a position only if ALL of the following are true:

1. Price threshold: Contract price is in $[0.78, 0.92]$. Below 0.78, the loss-to-win ratio may favor longshots. Above 0.92, the profit margin per win is too thin to absorb any miscalibration.

2. Fee-adjusted EV positive: Your model's estimated true probability exceeds the break-even probability (market price + fee drag) by at least 1.5 pp. If you cannot estimate true probability independently, you should not be trading.

3. Minimum remaining time: Do not enter a 5-minute contract with <90 seconds remaining, or a 15-minute contract with <4 minutes remaining. Late entries face thin books, wide spreads, and oracle uncertainty.

4. Order type preference: Always attempt a limit order (maker) first. Only fall back to taker (market/FOK) if the maker order has been resting for ≥15 seconds without fill and the opportunity is still valid. Zero-fee maker fills dramatically improve expected value.

5. Executable liquidity exists: Top-of-book ask depth is at least 5× your intended size within 2 pp of the current ask.docs.polymarket.com

6. Correlated exposure check: You have zero or one other open position in the same BTC-correlated group (see Rule 4).

7. No stale data: Market websocket is healthy, no recent sequence gaps, clock skew below internal threshold.

---

Rule 2: When NOT to Enter

Never enter under any of the following conditions, regardless of signal strength:

1. Late in window: Within the final 25% of the contract window (final 75s for 5-min; final 225s for 15-min). Non-negotiable. Last-second reversals are the single largest unmodeled risk.

2. High-volatility filter: Within 90 seconds of any external volatility shock (>0.5% BTC move in under 60 seconds on spot markets) or if the underlying coin has moved >1.5% in the last 5 minutes.

3. Thin order book: Top-of-book ask depth under $\$100$–$\$150$ for the target contract, or bid-ask spread exceeds 5 cents (or >1.5× its 1-hour rolling average).

4. Daily/weekly loss limit hit. See Rule 3.

5. Max correlated exposure already reached. See Rule 4.

6. Back-to-back losses: After 3 consecutive losing trades, pause for 2 full windows before the next entry.

7. During open position: Never enter a new position while a position in the same coin is already open.

8. API/data latency degraded: If your data feed has been delayed by >5 seconds in the last 60 seconds, skip the window entirely.

9. Post-shock macro events: Disable trading 5 minutes before and after major macro releases (FOMC, CPI, NFP) or scheduled network upgrades.

10. Rapid price decay: Do not enter if the contract price has dropped from 0.90→0.80 in under 10 seconds (indicates reversal, not consolidation).

---

Rule 3: Position Sizing

Use worst-case loss, not expected win rate:

$$\text{Position} = \min\left(\frac{1}{4} \times f_{\text{Kelly}} \times \text{Bankroll},\;\text{MaxPositionUSD}\right)$$

Where:

$$f_{\text{Kelly}} = \hat{p} - \frac{1 - \hat{p}}{b}, \quad b = \frac{\text{net win per share}}{\text{net loss per share}} = \frac{(1-p) - \text{fee}}{p + \text{fee}}$$

• Use quarter-Kelly (divide Kelly fraction by 4). Full Kelly is suicidal with estimated probabilities.
• Hard cap: No single trade may exceed 0.25% of NAV in worst-case loss, or an absolute dollar cap (start at $\$10$, graduate to $\$25$, then $\$50$ only after 200+ validated live trades).
• Daily loss limit: Stop all trading after losing 1.0% of NAV in a single calendar day.
• Weekly loss limit: If week-to-date losses exceed 2.0% of NAV, halt all trading and review.
• Never use fixed dollar sizing without a bankroll fraction cap. As bankroll shrinks, fixed sizing exponentially increases risk of ruin.

---

Rule 4: Max Correlated Exposure

At any point in time, total capital at risk across correlated assets must not exceed 1.5% of NAV.

Define correlation groups based on BTC beta:

• Group A (BTC-dominant, β > 0.80): BTC, ETH, SOL, BNB, AVAX — these move together in crashes
• Group B (lower beta): Altcoins with partial BTC correlation
• Group C (low beta): Truly idiosyncratic (rare in Polymarket crypto)

Hard rule: Maximum 1–2 open positions in Group A at any time. Period. The correlated blowup analysis shows holding 5 Group-A positions simultaneously raises $P(\text{all lose})$ to ~5.9% — unacceptable.

Aggregate exposure of all correlated directional trades must not exceed 3% of total bankroll.

Cross-window exposure: Positions in non-overlapping time windows (e.g., the 10:00 contract and the 10:15 contract) may be treated as independent, unless a macro event spans both windows.

If correlation spikes intraday, shrink caps automatically.

---

Rule 5: Multiple Entries on the Same Coin

Default: NO. One open position per coin per contract period. No averaging down. No pyramiding.

• If BTC-Up at 0.85 has moved to 0.80, adding more does not "improve" your entry — it doubles your exposure to the same binary outcome.
• Across consecutive contracts: Allow re-entry on the same coin in the next contract period only if the previous contract resolved and the new contract independently meets all entry criteria.
• Maximum 4 consecutive entries on the same coin in the same direction. After 4 consecutive wins in one direction, require a skip of at least 1 contract.
• If you later relax this: allow at most one add-on, only if it improves average price, only if total coin exposure stays within original trade-risk cap, and only after a cooldown.

---

Rule 6: When to Disable a Coin Entirely

Automatic disable if any of the following occur:

1. Consecutive loss trigger: 4–5 consecutive losses on that coin's contracts — disable for 24–48 hours and review.

2. Calibration drift: Rolling 20–100 trade win rate falls below the breakeven probability by more than 2 pp — disable until investigation is complete.

3. Liquidity collapse: Average bid-ask spread exceeds 8 cents over a 1-hour window, or top-of-book depth falls below $\$100$ for >30 minutes.

4. Regime change: Underlying coin's spot price shows a move >3% in any 5-minute window in the last 24 hours (volatility regime shift indicator). Disable for 12–24 hours.

5. Realized slippage exceeds modeled slippage by >50% over 10+ trades.

6. Oracle/resolution anomaly: Any contract on that coin resolved unexpectedly or was disputed.

7. Platform issues: Matching engine restart, API degradation, or order-status inconsistencies → disable all coins immediately until confirmed resolved.docs.polymarket.com

Re-enable criteria: Coin can be re-enabled after 48 hours of observation and after confirming win rate returns to above break-even in shadow mode for 20+ windows. Requires explicit human review.

---

Rule 7: Monitoring and Circuit Breakers

• Data freshness: If market price has not updated in 10+ seconds, assume data feed failure — do not trade.
• Execution confirmation: If order submitted and not confirmed within 8 seconds, treat as failed. Do not submit a second order without confirming the first is genuinely unconfirmed.docs.polymarket.com
• Position reconciliation: End of every window, reconcile your position log against on-chain state. Any discrepancy of >$\$5$ halts trading until resolved manually.
• Fee discrepancy alert: If realized fees differ from modeled fees by >20% on average over 10 trades, halt and re-verify the fee formula implementation.
• Request throttling detection: Official docs note throttling delays requests rather than rejecting — deadly for short-horizon bots. Monitor and alert.docs.polymarket.com

---

Rule 8: Logging and Audit

Every single decision (enter, skip, exit, disable) must be logged with:

• Timestamp (ms precision)
• Contract ID, coin, direction, time window
• Market price at decision time
• Model's estimated true probability
• Orderbook top-of-book bid/ask/depth at decision time
• Fee calculated
• Position size and rationale
• Fill price and fill latency (if entered)
• Trade status lifecycle (MATCHED → CONFIRMED or FAILED)
• Resolution outcome (once known)
• All concurrent open positions (for correlation audit)

Without this log, you cannot distinguish skill from luck, and you cannot debug failure modes. This is not optional infrastructure — it is the strategy.

---

Final Warning

The single most dangerous belief in short-horizon prediction market trading is:

"I win 85% of the time, therefore I have edge."

An 85% win rate at $p = 0.85$ with zero calibration edge produces negative expected value after fees. The fee is the rake, and the rake always wins over a sufficient number of hands. The only way to profit is to be provably better-calibrated than the market price, after accounting for fees, slippage, latency, and correlated exposure.

If your shadow-mode data does not show this edge at statistical significance, the correct action is to not trade. The market is not obligated to give you an edge just because you built a bot.

In these markets, treat execution as the strategy. Signal quality is secondary; cost, latency, and exposure discipline decide survival.