0x04b6d7

Wallet: 0x04b6d7e930cf9e493c5e6ef24b496294f95594c8 Window: 2026-04-22 to 2026-05-14 (23 calendar days, 20 active) Universe: 401,544 trades · 3,100 markets · $5.03M gross BUY turnover Net P/L: +$36,132 on $5,027,768 deployed = +0.72% ROI in 23 days

The Punchline

This is a high-frequency spread-capture bot with an embedded directional signal. The strategy is built around a simple mechanic: enter both sides of every short-duration crypto Up/Down market, lock in a combined cost below $1.00, and collect the $0.03-$0.05 spread guarantee when one side resolves at $1.00. The 98.2% both-sides participation rate (3,045 of 3,100 markets) makes the archetype unambiguous.

But "pure market maker" is too simple. The bot tilts its allocation based on a directional view: in 786 markets (25% of the book) it placed 3x+ more capital on one side than the other, and that dominant side won 92.9% of the time. In 650 markets (21%) it used a 2-3x tilt and the dominant side won 82.6%. The tilt accuracy is far too high to be random - there is a real signal driving the asymmetric allocation, and it is the most interesting part of this strategy.

The economics are disciplined but thin. $5.03M of BUY notional generates +$36,132 of net realized P/L. The spread P/L component computes to +$189,723 from guaranteed spread mechanics, but the losing-side hedge tax across all paired markets costs -$1,947,327. The directional accuracy on high-conviction tilts does most of the work recovering that tax. Remove the 3x+ conviction bucket and the book likely goes negative.

What He Trades

The universe is short-duration crypto Up/Down markets on Polymarket:

btc-updown-5m-*     (primary workhorse, ~55% of volume by trade count)
btc-updown-15m-*    (secondary)
eth-updown-5m-*     (significant, appears in early CSV alongside BTC)
eth-updown-15m-*    (minor)
sol-updown-5m-*     (visible in CSV tail, minor)
sol-updown-15m-*    (visible in CSV tail, minor)
xrp-updown-5m-*     (visible in CSV tail, minor)

The CSV sample from April 22 shows the bot simultaneously active in btc-updown-5m-1776839700, btc-updown-15m-1776839400, eth-updown-5m-1776839700, eth-updown-15m-1776839400, solana-up-or-down-april-22-2026-2am-et, and xrp-updown-5m-1776839700 within the same 3-minute window. This is a multi-asset sweeper that enters every available crypto Up/Down market as it opens.

The category breakdown confirms the near-total concentration:

The "Other" category is 33 trades totaling $453 - rounding error. The entire book is crypto Up/Down.

The Order of Operations - One Market, Trade by Trade

The CSV sample covers a dense 2-minute window at the open of btc-updown-5m-1776839700 (Bitcoin Up or Down - April 22, 2:35AM-2:40AM ET). This market resolved "Up" and the bot's operation is fully visible.

Walk-through:

1. 06:37:08 - Market opens. The bot fires a burst of BUY Up orders at $0.52 (roughly fair value for a 50/50 event). Four fills in the same second, totaling ~$52. This is the initial "Up" position at mid-market.

1. 06:37:10-06:37:18 - Continuing to build Up at $0.48-$0.49. The bot walks the book down as it exhausts the best asks. The Up VWAP across this phase is approximately $0.49.

1. 06:37:18-06:37:34 - The bot simultaneously enters Down at $0.47-$0.51. This is the paired entry. While still buying Up, it opens the Down side. Within 16 seconds of its first trade, both sides are active. The Down side receives significantly more capital than Up in this window: roughly $120 into Down vs $65 into Up so far.

1. 06:38:26-06:38:34 - Late Down entries at $0.78-$0.84. The bot adds more Down shares at much higher prices, suggesting the market has moved strongly toward "Up" and Down is now pricing as a longshot. This is the "hedge reinforcement" pattern - adding cheap Down shares at high probability prices late in the window to improve the non-dominant side's position for spread purposes.

1. 06:38:32-06:38:40 - Cheap Up cleanup at $0.14-$0.18. Now that Up is pricing near-certain, the bot sweeps cheap Up asks to ensure it has coverage on both sides at favorable blended prices.

Resolution: "Up" wins at 2:40AM. The Up shares pay $1.00. The Down shares pay $0.00. Given the much larger Down allocation at $0.47-$0.84, this market was a net loss on the directional component - but the spread (the gap between the Up VWAP + Down VWAP vs $1.00) still delivers a guaranteed slice regardless of outcome.

The critical observation from this specific market: the bot showed a Down tilt (more capital into Down) but "Up" resolved. This is not a failure of the strategy - the paired structure guarantees a positive expected value from the spread, and the directional signal only needs to be right across the population. The aggregate 92.9% accuracy on 3x+ tilt markets confirms the signal works across the book.

Why It Works - The Math

The spread-capture mechanic is guaranteed by construction:

For a paired market:
  Up VWAP   = sum(up_usdc) / sum(up_shares)
  Down VWAP = sum(down_usdc) / sum(down_shares)
  Paired cost = Up VWAP + Down VWAP

  If paired cost < 1.00:
    One side always resolves at $1.00 (Up or Down, never both)
    Guaranteed profit per paired share = 1.00 - paired_cost

Median paired cost observed: $0.9679
Guaranteed profit per $1.00 of paired shares: $0.0321 (3.2%)

For 3,045 markets with a median paired cost of $0.9679, the structural spread profit is +$189,723 before the hedge tax. The hedge tax is the losing side's total cost: $1,947,327. The bot recovers this through its directional accuracy on tilted allocations.

High-conviction P/L decomposition (3x+ bucket, 786 markets):
  Dominant-side win rate: 92.9%
  Markets won: ~730 of 786
  When dominant side wins: dominant-side payout at $1/share
                           non-dominant (hedge) side: total loss
  When dominant side loses: non-dominant pays $1/share
                            dominant side: total loss
  
  EV per market (simplified):
    Win (92.9%):  dominant_shares × $1.00 - dominant_cost - hedge_cost
    Loss (7.1%):  hedge_shares × $1.00 - dominant_cost - hedge_cost
    
  With 3x dominance ratio and average paired cost 0.98:
    Win EV = 0.75 × $1.00 - $0.75 - $0.25 + spread gain
    = +$0.03+ per paired dollar (from spread) + directional upside

The strategy's EV is positive across the population because: (a) the spread guarantee provides a floor, and (b) the directional signal on high-conviction markets generates substantial directional profit above the spread floor.

---

Phase 1 - Trader Profile

Scale and Activity:

This is a SELL-free book. Zero sell transactions. The bot holds every position to settlement. This is architecturally consistent with a spread-capture strategy where the payout at resolution ($1.00 win / $0.00 loss) is the exit mechanism. There is no SELL engine - positions are entered and held until the market closes.

Trade Size Distribution:

The distribution is mildly top-heavy but not power-law. The max ($12,221) is an outlier - one single fill into btc-updown-5m on May 10 that resolved as a win (+$123). The P99 of $124 and P95 of $60 show that the vast majority of the book is small clips ($3-$60). The Lorenz curve shows the bottom 50% of trades carry only 5.9% of capital, and the top 10% carry 58%. Moderate concentration, consistent with a bot that sizes up on high-conviction tilts.

Execution Signature:

Median 0.0 seconds and 87% of fills under 10 seconds: pure bot execution. The CSV confirms simultaneous multi-market operation - at 06:37:08 UTC on April 22, fills fire across btc-5m, eth-5m, btc-15m, sol-hourly, and xrp-5m within seconds of each other. No human operator can manage this. The 7-second mean with 0-second median reflects bursts of same-second fills (market entry) interspersed with brief inter-market pauses.

Trading Hours (UTC):

The sleep window is 21:00-03:00 UTC - a 6-hour overnight gap. Combined with the US session concentration, this is likely a US-East-Coast operated system running 4am to 4pm Eastern time.

Archetype:

SPREAD CAPTURE + DIRECTIONAL TILT

Both-sides spread capture as the structural foundation, with an embedded directional signal that skews allocation toward the expected winner, particularly at 3x+ conviction levels.

---

Phase 2 - Core Strategy Identification

Both-sides participation: 98.2%

3,045 of 3,100 markets had both Up and Down sides purchased. The 55 one-sided markets (1.8%) are likely cases where the bot's second-side order didn't fill before resolution, not intentional one-sided bets.

Both-sides statistics:

A median paired cost of $0.9679 means the bot successfully locks in a 3.2-cent spread guarantee on the typical market. 64.6% of paired markets have a cost below $1.00, meaning the bot successfully achieves the spread-positive threshold on nearly two-thirds of its entries. The remaining 35.4% have paired costs above $1.00 - these markets are net-negative from the spread mechanic and must be covered by directional accuracy.

Second-side lag: median 16 seconds.

The bot enters both sides within 16 seconds on the typical market. This is functionally simultaneous - both sides are established well before any price movement could make the second entry adversarial. This is paired from inception, not reactive hedging.

Classification: This is archetype A (Both-Sides Spread Capture) with strong elements of archetype B (Directional Betting) on high-conviction tilted markets. The bot buys both sides on 98.2% of markets (A) but tilts allocation based on a directional signal with measurable accuracy (B).

---

Phase 3 - Dominance Ratio Analysis

This is the most important analytical section. The dominance curve reveals a clean, monotonic relationship between allocation tilt and directional accuracy:

The jump in dominant-side win rate from 55.6% at 1.0-1.5x to 92.9% at 3x+ is the signal. This monotonic escalation is impossible to explain by chance or by selection bias in paired-cost quality. The bot has a real directional edge that it expresses through asymmetric allocation: small position on the "hedge" side, large position on the expected winner.

The 3x+ bucket's 92.9% accuracy on 786 markets is the single most important number in this report. It means 730 of 786 high-conviction calls resolved correctly. The mean paired cost in the 3x+ bucket ($0.9798) is slightly higher than the 1.0-1.5x bucket ($0.9444) - meaning the bot doesn't sacrifice spread quality to express conviction, it just sizes up aggressively on the favored side.

Implications for strategy classification: The 1.0-1.5x bucket (1,002 markets, 55.6% WR) is pure spread capture - the bot has no strong view and buys near-equal amounts on both sides, collecting the spread guarantee. The 3x+ bucket (786 markets, 92.9% WR) is functionally a directional bet with a token hedge to lock in the spread. Both mechanics coexist in the same wallet.

---

Phase 4 - Entry Price Analysis

The win rate column is a near-perfect calibration of the prediction market. 5.9% wins at sub-$0.10 (the market prices these as ~5-10% probability events), 96.1% at sub-$1.00 (the market prices these as ~95%+ probability events). The market is pricing outcomes correctly, and the bot enters across the full price spectrum as a natural consequence of buying both sides.

The ROI column shows the actual alpha locations. The $0.30-$0.40 band (+2.4% ROI) and $0.60-$0.80 bands (+2.2% ROI each) are the sweet spots. The $0.20-$0.30 band is the worst performer at -4.0% ROI on $321K of capital - a notable drag. The $0.50-$0.60 band is also slightly negative (-1.0%).

The two largest capital bands are $0.40-$0.50 and $0.50-$0.60 (each ~$740K), which together hold 29.5% of total capital but produce only +$4,466 and -$7,390 respectively. The coin-flip zone is where the spread guarantee is most valuable (both sides near $0.50 = lowest paired cost achievable) but also where directional alpha is hardest to extract.

Sub-bucket observation: The price distribution across the full book spans all 100 cents from $0.01 to $0.99. There is no single-tick concentration. The bot enters wherever the book offers supply - it is not pegging to a specific fair-value level, it is walking the entire depth of both sides simultaneously.

---

Phase 5 - Category and Market-Type Breakdown

Single category (Crypto, +0.72% ROI, +$36,323 P/L). The "Other" bucket is 33 trades totaling $453 - negligible and likely misclassified crypto slugs.

Market type breakdown from CSV:

The top markets by volume confirm BTC dominance: every entry in both the top-10 by volume and top-10 by P/L is a Bitcoin Up or Down market. ETH, SOL, and XRP appear to be lower-volume satellites where the bot allocates smaller clips.

Best and worst single-market outcomes:

The best market: "Bitcoin Up or Down - May 1, 12:00PM-12:15PM ET" - 410 trades, $6,356 volume, +$2,552 P/L, 75.9% WR. This is a strongly tilted market where the bot had a high-conviction call that paid off.

The worst market: "Bitcoin Up or Down - May 1, 11:30AM-11:45AM ET" - 274 trades, $5,448 volume, -$3,400 P/L, 8.4% WR. Near-total loss: 251 of 274 trades resolved against the dominant side. This is the hedge tax in action - the bot had a large dominant-side bet that resolved opposite to expectations.

---

Phase 6 - Timing and Execution Analysis

Hourly P/L and trade distribution:

The 4 worst hours by P/L (identified in filters as hours 4, 7, 9, 14) collectively drag -$4,544. Hour 16 UTC alone (+$13,089) exceeds the entire book's net P/L.

The peak P/L hours - 13:00-18:00 UTC (9am-2pm Eastern) with a standout at 16:00 UTC (12pm Eastern) - align with peak BTC/ETH spot volatility during the US trading session. More volatile spot markets likely generate more CLOB mispricings, enabling better entry prices on both sides and more frequent 3x+ conviction opportunities.

Day-of-week performance:

Wednesday is the dominant day by a large margin: +$26,125 of P/L is 72% of the entire 23-day net P/L. Monday is the worst full-volume day at -$3,918. The Saturday sample (135 trades) is too small to draw conclusions.

Entry timing within the market window:

From the CSV sample, the bot begins entering within 2-8 seconds of a market opening and typically completes its full book-walk within 60-120 seconds. It does not accumulate throughout the window - it front-loads entry to secure the best (lowest) prices before competition arrives. This is consistent with a bot that fires on market-open events.

---

Phase 7 - Filter Experiments

Full filter analysis in the companion filters tab. Summary table:

The high-conviction filter is transformative: restricting to only the dominant side of 2x+ markets converts a 0.72% ROI book into an 18.05% ROI book. This is the single most important finding from the filter analysis. The low-conviction, near-equal markets (1.0-1.5x bucket) are diluting the return on the high-conviction signal substantially.

---

Phase 8 - Rolling Window Consistency

Weekly summary:

Rolling 7-day window analysis:

• Windows positive: 17 of 23 (73.9%)
• Windows negative: 6 of 23 (26.1%)
• Peak 7-day P/L: +$19,149 (week ending Apr 26)
• Trough 7-day P/L: -$6,436 (week ending May 1)

Rolling 15-day window analysis:

• All 23 rolling 15-day windows: positive
• Range: +$5,686 (May 11) to +$22,079 (May 14)
• 100% of 15-day windows are green

The 7-day consistency is lower than SirMartingale's 100% green rate - the strategy has a negative week (W18, -$6,436) driven by the catastrophic May 1 markets (the two worst markets in the book both occurred on May 1 in the 11:15-11:45AM window, totaling -$6,448 in just 573 trades). This is the strategy's vulnerability: a cluster of high-conviction incorrect calls in a short time window can erase weeks of spread-capture gains.

---

Phase 9 - P/L Decomposition

The decomposition reveals something unexpected: the spread guarantee alone (-$1.76M after hedge tax) is net negative. The strategy only profits because the directional signal on high-conviction tilted markets generates dominant-side wins that more than cover the hedge tax. The spread mechanic is not the source of profit - it is a partial cost recovery mechanism that reduces the effective cost of being wrong. Directional accuracy is the real edge.

This reframes the archetype: the bot is best described as a directional bettor that uses spread-capture mechanics to reduce downside risk, not a market maker that occasionally tilts.

---

Phase 10 - Strategy Specification

One-sentence summary: A high-frequency bot that simultaneously buys both sides of every short-duration crypto Up/Down market on Polymarket, using a directional signal to skew allocation toward the expected winner (3x+ on high-conviction markets), collecting the spread guarantee as a floor and directional P/L as the ceiling.

Edge source: Two stacked mechanisms: (1) paired-cost spread guarantee providing a P/L floor independent of outcome, and (2) a directional signal (likely spot-price-based fair-value model) generating 92.9% accuracy on high-conviction (3x+) market allocations.

What works: The 3x+ dominance bucket (786 markets, 92.9% WR) is the alpha engine. Excluding it, the book would be approximately break-even or negative. The 16:00-18:00 UTC window (+$21K over the window) is the daily performance peak. Wednesdays drive 72% of net P/L.

What drags: The low-conviction 1.0-1.5x bucket (1,002 markets) generates minimal P/L while deploying significant capital. Hour 09 UTC (-$2,012) and the May 1 cluster (-$6,448 in 30 minutes) represent the strategy's left-tail risk. The $0.20-$0.30 price band shows -4.0% ROI on $321K of capital and is the single worst price zone.

Full spec in Playbook tab.

Wallet: 0x04b6d7e930cf9e493c5e6ef24b496294f95594c8 Window: 2026-04-22 → 2026-05-14 (20 active / 23 calendar days) Methodology: Cash-flow P/L = -buy_usdc + sell_usdc + remaining_share_payout. Resolved shares settle at $1 (win) / $0 (loss); open positions marked at last price.

Phase 1 - Trader Profile

Scale

Trade-size distribution (USDC per fill)

Inter-trade gap, same (market, outcome)

Phase 2 & 3 - Both-Sides Participation, Dominance Curve

• Both-sides rate: 98.23% (3,045 of 3,100 markets)
• Median paired cost: $0.9679
• Mean paired cost: $0.9644
• Paired cost % under $1.00: 64.6%
• Paired cost % under $0.97: 51.2%
• Median 2nd-side hedge lag: 16s

Dominance buckets

Phase 4 - Entry-Price Analysis

Phase 5 - Category & Vertical Breakdown

Phase 6 - Timing & Execution

Net P/L by hour (UTC)

Phase 8 - Rolling Window Consistency

• Rolling 7-day windows green: 19 of 23 (82.6%)
• Rolling 7-day P/L range: -$6,436 → +$19,149
• Rolling 15-day windows green: 23 of 23 (100.0%)
• Rolling 15-day P/L range: +$5,686 → +$22,079

Weekly P/L

Phase 9 - P/L Decomposition

Phase 10 - Top Markets by Volume

Top 10 winners by P/L

Top 10 losers by P/L

Report generated 2026-05-15 07:21 UTC.

Wallet: 0x04b6d7e930cf9e493c5e6ef24b496294f95594c8 Window: 2026-04-22 to 2026-05-14 Baseline: 401,544 BUYs · 42.8% WR · $5,027,768 deployed · +$36,132 P/L · +0.72% ROI

---

The headline result

One filter transforms this from a 0.72% ROI book into an 18% ROI book. The others are marginal adjustments.

The high-conviction filter (dominant side only, dominance ratio 2x+) is the single filter that matters. It cuts 83% of the trade count but extracts +$309K P/L at +18.05% ROI - versus +$36K at +0.72% unfiltered. This is not a marginal improvement; it is a strategy-defining finding. The low-conviction markets (1.0-1.5x paired buys) are diluting the book heavily.

Every other filter - price band, category, hour exclusion - produces single-digit percentage improvements that don't change the strategic picture. The combined filter (stacking price + dominance + hours) underperforms the dominance-alone filter because it adds price-band restrictions that cut good high-conviction markets alongside bad ones.

---

Filter results table

---

Filter-by-filter commentary

1. Price band filter ($0.30-$0.70) → MODEST LIFT

The $0.30-$0.70 filter removes the extreme longshot zone ($0.00-$0.30) and the near-certain zone ($0.70-$1.00), keeping only the coin-flip-area trades. Result: 221,928 trades, 47.1% WR, +$27,495 P/L, +1.02% ROI.

That is a +30bp improvement in ROI (+1.02% vs +0.72%) at the cost of cutting the book to 55% of its original size. In absolute dollars, the filter loses $8,637 compared to baseline.

For a directional bettor, this filter typically concentrates the book on the best-calibrated odds range. For this paired spread-capture bot, the interpretation is different. The $0.20-$0.30 band is genuinely bad: -4.0% ROI on $321K. The $0.50-$0.60 band is slightly negative (-1.0% on $744K). Removing those does help. But the $0.70-$0.80 and $0.90-$1.00 bands both carry positive ROI and $1.35M of capital - and this filter discards them.

Verdict: Modest improvement in percentage ROI but absolute P/L degrades. Not worth applying if you are capacity-constrained on bankroll. Might be worth applying if you are running a smaller book where capital efficiency matters more than absolute P/L.

2. High-conviction filter (dom ≥ 2x, dominant side only) → MEANINGFUL_LIFT

This is the only filter that materially changes the strategy. Restricting to the dominant-leg-only of markets where the bot allocated at least 2x more to one side than the other produces:

• 69,886 trades (17.4% of baseline)
• 83.7% win rate (vs 42.8% baseline)
• $1,712,145 capital (34% of baseline)
• +$309,046 P/L on a +18.05% ROI

The +$272,914 delta over baseline is achieved by throwing away 83% of the book. The implication is stark: the low-conviction paired buys are collectively destroying value. Those 1.0-1.5x markets - 1,002 of them - add volume and spread guarantee but their net directional P/L is negative or marginal. The 2x+ markets are where the real edge lives.

The further refinement from the dominance bucketing shows:

A replicator who could access only the 3x+ bucket would have an even cleaner book: 786 markets, ~93% win rate, near-directional-only economics.

Verdict: The single most impactful filter available. A replicator implementing only the high-conviction dominant-leg signal would run an 18% ROI book versus the baseline 0.72%. The trade-off is capacity: only 17% of the original trade count qualifies.

3. Category filter (Crypto only) → NO-OP

The full book is 99.99% Crypto by volume. The "Other" category (33 trades, $453 volume, -$190 P/L) is statistically irrelevant. Applying the Crypto-only filter improves P/L by exactly $191, which is the "Other" category loss recovered.

The category filter is a formality here. There is nothing to filter. The bot already restricts itself to crypto Up/Down markets by design.

4. Hour exclusion filter (exclude hours 4, 7, 9, 14 UTC) → MODEST LIFT

The four identified worst hours are:

Excluding these four hours removes 74,087 trades and improves P/L from +$36,132 to +$37,394 - a +$1,262 improvement. ROI improves from +0.72% to +0.92%.

However, hour 14 UTC (9am Eastern, US session open) is technically listed in the worst-hours set but shows +$3,282 P/L - it is positive, just lower-ranked. Excluding it actually removes good trades. The real drag hours are 04:00, 07:00, 09:00, and the genuinely bad ones below 13:00 UTC.

Verdict: A +$1,262 absolute improvement over 23 days is real but modest - $55/day. The filtering mechanics here are weak because many negative-P/L hours have large trade counts that include good trades alongside bad ones. The hour filter is a blunt instrument on a book with this many trades per hour.

5. Combined filter (price + dominance + hours) → WORSE THAN DOMINANCE ALONE

The combined filter produces: 181,632 trades, 47.5% WR, $2.16M capital, +$23,040 P/L, +1.06% ROI.

Compared to the high-conviction filter alone (+$309,046 P/L, +18.05% ROI), the combined filter is dramatically worse in both absolute P/L and ROI. Adding the price band and hour restrictions to the dominance filter cuts $285K of profit. This happens because:

1. The price band filter removes high-conviction markets that happen to occur at extreme price points (sub-$0.30 or above $0.70), which are valid opportunities when the directional signal is strong.
2. The hour filter removes some high-conviction markets that fall in the "bad" hours (like hour 14 which is actually P/L positive).

Verdict: Do not stack these filters. The dominance filter alone is optimal. Adding price or hour constraints on top of it reduces the qualifying set without proportional reduction in bad trades.

6. A filter the standard battery does not test: worst-day exclusion

The May 1, 11:15AM-11:45AM ET cluster caused -$6,448 of P/L in 573 trades. If a circuit-breaker filter paused trading after -$3,000 of daily P/L, Week 18 (-$6,436 weekly P/L) would have been approximately break-even instead of the book's only negative week.

This is not a standard PR&R filter but it is the most practically useful risk management tool for this strategy. A daily drawdown limit of -$3,000 to -$5,000 would have preserved the Week 18 P/L without meaningfully clipping the upside weeks.

---

Bottom line for replication

Three actionable findings from this filter analysis:

1. Apply the high-conviction filter. The dominant side only, for markets with dominance ratio 2x+. This is the difference between a 0.72% ROI grind and an 18% ROI strategy. The capacity reduction (83% fewer trades) is acceptable if your directional signal can identify the 3x+ conviction opportunities.

2. Avoid the $0.20-$0.30 price zone. This band costs -$12,703 on $321K deployed (-4.0% ROI) and is the single worst segment of the book. If you cannot filter it naturally through your entry logic, add an explicit exclusion for entries priced $0.20-$0.30.

3. Consider a daily drawdown circuit-breaker. The May 1 cluster showed how a sequence of incorrect high-conviction calls can erase a week of gains in 30 minutes. A -$3,000 to -$5,000 daily loss limit that pauses trading for the remainder of the day would protect against the left-tail event without meaningfully reducing expected profit.

Source wallet: 0x04b6d7e930cf9e493c5e6ef24b496294f95594c8 Strategy: Both-sides spread capture with directional tilt on crypto Up/Down markets Reference book: $5,027,768 BUY notional · +$36,132 net P/L · +0.72% ROI (full book); +18.05% ROI (high-conviction filter only)

---

One-paragraph operator brief

Build a Polymarket bot that simultaneously enters both sides of every BTC, ETH, SOL, and XRP Up/Down market (5-minute, 15-minute, and hourly windows) within seconds of market open. Walk the full depth of both sides to achieve a combined entry cost below $0.97. In parallel, compute a fair-value estimate from the spot price and allocate asymmetrically: if one side is mispriced relative to fair value by more than your threshold, put 3x+ capital on the favored side and 1x on the hedge side. Hold all positions to settlement. The spread guarantee provides a floor; the directional signal at 3x+ tilt provides the real P/L. Run 05:00-20:00 UTC; sleep 21:00-04:00 UTC. Do not apply a price-band filter to the entry signal - the directional accuracy justifies entries across the full price spectrum at high-conviction moments.

---

1. Market selection

Why these markets only: Short-duration crypto Up/Down markets have predictable resolution windows (5-15 minutes), small capital requirements per market, and a spot-price proxy (BTC/ETH/SOL/XRP tick data) that enables fair-value computation. The 3,100 markets across 23 days implies approximately 135 markets open per active day - enough throughput to absorb the full working capital without over-concentrating in any one market.

Eligibility gates:

• Market must be live and not within 30 seconds of close when bot first enters
• Both Up and Down sides must have at least $50 of available depth in the $0.10-$0.90 price range
• Skip markets already 120+ seconds old at the moment the bot first processes them (stale entry disadvantage)

---

2. Entry logic

The bot runs two parallel entry engines simultaneously: the paired sweep and the directional tilt.

Paired sweep (fires on every eligible market):

def paired_sweep(market, budget):
    """Walk both sides of the orderbook to achieve paired cost < 0.97"""
    up_budget   = budget * 0.5     # start equal
    down_budget = budget * 0.5
    
    # Apply directional tilt BEFORE calling this - see Section 3
    up_fills   = walk_book_buy(market, "Up", max_usdc=up_budget)
    down_fills = walk_book_buy(market, "Down", max_usdc=down_budget)
    
    up_vwap   = up_fills.usdc   / up_fills.shares
    down_vwap = down_fills.usdc / down_fills.shares
    paired_cost = up_vwap + down_vwap
    
    if paired_cost > 1.05:
        # Over-paid - do not continue adding; accept current position
        log_warning(f"Paired cost {paired_cost:.4f} on {market.slug}")
    
    return up_fills, down_fills

Entry price target:

• Target Up VWAP: $0.46-$0.50 (slightly below mid)
• Target Down VWAP: $0.46-$0.50 (slightly below mid)
• Combined target: $0.92-$0.98
• Hard exit if paired cost exceeds $1.05

Timing within the window:

• Enter within the first 60 seconds of market open
• Complete the paired entry within 120 seconds
• No re-entry to add shares after minute 2 - the book is set

---

3. Directional tilt (the alpha layer)

This is the component that elevates ROI from 0.72% to 18%+ on the qualified subset. The bot must have a real-time fair-value model.

def compute_directional_tilt(market, spot_price_now):
    """
    Returns (tilt_side, tilt_ratio) where tilt_ratio is the
    multiplier applied to the favored side's budget.
    tilt_ratio = 1.0 means no tilt (equal allocation).
    tilt_ratio = 3.0 means put 3x on favored side vs hedge side.
    """
    fair_prob_up = spot_to_fair_prob(
        asset     = market.asset,
        threshold = market.threshold,
        spot      = spot_price_now,
        vol       = realized_vol_1h(market.asset),
        ttl_secs  = seconds_until_close(market)
    )
    
    clob_mid_up   = market.up_side.mid_price
    signal_gap    = fair_prob_up - clob_mid_up
    
    if abs(signal_gap) < 0.03:
        return None, 1.0          # no conviction: equal allocation
    elif abs(signal_gap) < 0.06:
        tilt_ratio = 1.5          # modest tilt
    elif abs(signal_gap) < 0.10:
        tilt_ratio = 2.0          # moderate tilt
    elif abs(signal_gap) < 0.15:
        tilt_ratio = 3.0          # high conviction
    else:
        tilt_ratio = 4.0          # maximum conviction (≥ 3x+ bucket)
    
    favored_side = "Up" if signal_gap > 0 else "Down"
    return favored_side, tilt_ratio

def apply_tilt(budget, favored_side, tilt_ratio):
    """
    Given a total budget, split into favored/hedge allocations.
    tilt_ratio=3.0 → 75% favored, 25% hedge
    tilt_ratio=2.0 → 67% favored, 33% hedge
    """
    favored_share = tilt_ratio / (tilt_ratio + 1.0)
    hedge_share   = 1.0 - favored_share
    
    return {
        favored_side:  budget * favored_share,
        other_side:    budget * hedge_share
    }

Threshold calibration (from reference data):

---

4. Sizing model

The sizing framework operates at two levels: per-market budget and clip size within the market.

Per-market budget:

At $100K+ bankroll, the per-market clips become large relative to the typical 5-minute market's orderbook depth. Walking the book to fill $300-$500 per side starts to move your own prices unfavorably. Fragment across more simultaneous markets rather than increasing per-market budget.

Clip size within a market:

The bot walks the book in small clips, not one large order. From the CSV: individual fills range from $0.04 to $115 within a single 5-minute market. The effective fill algorithm walks the ask side in clips of $1-$20 each, accumulating shares until the per-market budget is exhausted.

def walk_book_buy(market, outcome, max_usdc):
    """Walk the orderbook in small clips until budget exhausted or book too thin"""
    spent = 0
    fills = []
    
    for ask_price, ask_volume in market.orderbook[outcome].asks:
        if spent >= max_usdc:
            break
        if ask_price > 0.97:      # stop buying expensive shares
            break
        
        clip_usdc  = min(ask_volume * ask_price, max_usdc - spent, 20.0)
        clip_shares = clip_usdc / ask_price
        
        fills.append(Fill(outcome, ask_price, clip_shares, clip_usdc))
        spent += clip_usdc
    
    return FillSummary(fills)

The $12,221 single fill outlier: The reference book shows one fill on May 10 at $12,221 into a single BTC 5m market. This is a 100x+ outlier versus the P99 of $124. It is likely a manual or test trade, not part of the bot's normal operation. Do not build this into the sizing model.

---

5. Paired allocation mechanics

Target paired cost:

Second-side lag:

The reference book enters both sides within a median 16 seconds. This is the target window. If the second side cannot be entered within 60 seconds (due to depth or latency), abort the second-side entry and accept a one-sided position. The one-sided position loses the spread guarantee but retains the directional component - which at 3x+ tilt may still be net positive.

Both-sides rate management:

The reference wallet achieves 98.2% both-sides rate. Target 95%+ in replication. The 1.8% one-sided markets likely represent cases where the second-side entry failed due to a completely depleted orderbook or near-resolution timing. Accept these as natural slippage.

---

6. Exit strategy

There are no active exits. The bot holds 100% of positions to settlement.

# Exit strategy:
# Hold all positions until market resolves.
# Winning side pays $1.00/share at resolution.
# Losing side pays $0.00/share at resolution.
# No sell orders, no early exit, no take-profit.

def on_market_resolve(market, winning_side):
    up_payout   = up_shares   × 1.00 if winning_side == "Up"   else 0
    down_payout = down_shares × 1.00 if winning_side == "Down" else 0
    net_payout  = up_payout + down_payout
    net_pnl     = net_payout - total_usdc_spent

Why no SELL engine? The spread-capture strategy depends on buying both sides at combined cost below $1.00. An active SELL engine would require selling shares back into thin books at unfavorable prices, negating the spread guarantee. The resolution payout of $1.00 on the winning side is the cleanest exit for paired positions.

Contrast with SirMartingale: SirMartingale's SELL engine is its alpha source. This wallet has no SELL engine. These are different strategies with different exit mechanics. Do not graft a SELL engine onto this paired-structure strategy.

---

7. Risk management

Per-market maximum:

• Hard cap: $500 per market at $25K bankroll scale
• The reference wallet's median market spend is ~$1,600 (total volume / 3,100 markets), but this includes both sides. Actual per-side exposure is ~$800 median.

Daily circuit-breaker:

The May 1 cluster (-$6,448 in 30 minutes) demonstrates the need for a daily loss limit:

DAILY_LOSS_LIMIT = bankroll * 0.12     # 12% of bankroll

def before_entering_new_market():
    if daily_pnl_today < -DAILY_LOSS_LIMIT:
        log("Daily loss limit hit - pausing until tomorrow")
        return SKIP
    return PROCEED

At $25K bankroll this is a -$3,000 daily limit. At $10K bankroll: -$1,200. This would have stopped the May 1 bleeding at approximately -$3,400 instead of -$6,448, preserving an otherwise negative week.

Concentration limits:

• Do not put more than 20% of bankroll into markets open simultaneously
• If 10+ markets are open concurrently, cap per-market budget at bankroll/50 to prevent over-concentration

Paired-cost monitoring:

Track the rolling 100-market median paired cost. If it rises above $0.99 (sustained, not one-off), the market is becoming more competitive and the spread guarantee is degrading. Reduce position sizes by 50% when median paired cost exceeds $0.99.

---

8. Hour scheduling

Hour exclusions (the four worst):

The filter analysis identifies hours 4, 7, 9, and 14 UTC as the "worst hours." However, hour 14 UTC (+$3,282 P/L) is actually positive - it appears in the worst list because its P/L is lower than neighboring hours, not because it is negative. The genuinely problematic hours are 07:00 and 09:00 UTC (both significantly negative). Consider a 50%-size reduction at hours 07:00-09:00 UTC rather than outright exclusion.

Day-of-week:

Do not shut down on Wednesdays for any reason - +$26,125 is 72% of total P/L and occurred on the 4 Wednesdays in the observation window. If operational maintenance is needed, schedule it for Saturday (135 trades in 23 days; appears nearly unused by the reference bot).

---

9. Operational requirements

---

10. Diagnostic checklist for "is the bot still working?"

Run weekly:

---

What this playbook deliberately does NOT include

No active SELL engine. Every position resolves at $1.00 or $0.00. Adding sells would require timing the exit against thin 5-minute orderbooks and would likely degrade the spread guarantee. The reference wallet has zero SELL trades and achieves its P/L entirely through settlement payouts.

No price band restriction on entries. The filter analysis shows the $0.30-$0.70 band filter produces modest ROI improvement at the cost of absolute P/L. On a high-conviction (3x+) signal, entries in the $0.10-$0.30 zone and $0.70-$0.90 zone are profitable. The directional accuracy justifies paying for shares that look like "longshots" or "near-certain" when the fair-value model says otherwise.

No diversification into non-crypto markets. The "Other" category (33 trades, -$190 P/L) confirms that non-crypto markets are not part of this strategy. The fair-value model requires a spot price proxy, which only exists for crypto assets.

No hold-time management. Some systems try to exit before resolution if the position is deeply in-the-money. This wallet holds to settlement. The 5-15 minute resolution windows are short enough that the opportunity cost of holding is negligible, and early exits require selling into thin books.

No scaling without fragmentation. Above $30K working capital, increase throughput by running multiple wallets concurrently, not by increasing per-market clip sizes. Large clips in 5-minute markets with limited depth will move prices against the entry and degrade the paired cost.