Hey,

Here's what I found this week.

Finding #1: Funding rates are not a free lunch — fees eat the edge alive BTC perpetual funding averages 3.99% APY. Sounds like easy carry money. It's not: Polymarket's 3% taker fee is 40× larger than the per-trade VRP. Funding rate arbitrage only works with maker-only orders on zero-fee venues, or on much longer holding windows than 5-minute binary options. Lesson: Always model fees before modeling edge.

Finding #2: The multi-factor filter turns 50/50 into 57% — but n=14 isn't enough to trust it Combining three signals (volatility regime, implied VRP, liquidity cluster proximity) produced 57.1% win rate and +0.12% net edge per trade on 30 days of real BTC data. Sounds like progress. The noise term (±0.15%) is larger than the signal. I need 100+ trades before I can believe this isn't variance. Lesson: Edge is easy to imagine, hard to confirm.

Finding #3: Polymarket dropped fees to 0% — then the bot ran and ACCEPTED Polymarket quietly dropped taker fees to 0% mid-week. That changed the strategy math overnight: the edge I calculated as -1.38% net (unprofitable) flipped to +0.12% net (viable). So I built a live paper trading bot and ran it against real Polymarket feeds.

The result: SPRT ACCEPTED at 22:24 IST tonight. - n=28 trades | 25W/3L | 89.3% win rate - Starting balance: $10.00 → Final balance: $47.75 (+377.5%) - Sequential Probability Ratio Test logLR: 2.823 (ACCEPT boundary: 2.773) - Assets traded: BTC, ETH, SOL, XRP (15-min markets, NO direction, mean reversion)

The SPRT formally validates the strategy with statistical confidence — not just a lucky streak. 28 trades to ACCEPT vs. the projected 120 for a fixed-sample test. Signal filtering accelerated convergence.

This is the most significant result of the project so far. Theory → backtest → live validation: done.

Finding #4: 84% of signals get skipped — and that's the edge The 89.3% win rate wasn't luck. It was the filter. A 3-gate system screens every generated signal before a trade fires: Gate 1 checks if the volatility regime is active (EMA crossover), Gate 2 requires estimated win probability ≥65%, Gate 3 confirms Kelly fraction is positive. Result: ~172 raw signals → only 28 trades. One more discovery: 5-minute Polymarket pools are effectively dead — near-zero liquidity makes them untradeable. 15-minute pools only. Lesson: Selectivity IS the edge. Most bots scan everything. This bot skips 84% of what it sees.

What didn't work: Pure volatility-selling is dead even at 0% fees — the raw VRP per trade (~0.037%) is still too small to build around. The only viable path is multi-factor stacking (regime + cluster + VRP together), not any single signal alone.

This week on the blog: → The Moment of Truth: Backtesting on Real BTC Data → Everything Just Changed: Polymarket 0% Fees & the Paper Trading Bot → The Kelly Criterion for Binary Options — How Much to Bet When You Have an Edge → Why I Skip 84% of Trades (Signal Filtering Deep Dive) → Paper Run 2: The Selectivity vs. Coverage Tradeoff → Day 11: The Dry Run That Saved $10.49 → Day 12: The Fee Flip — From Paying 10% to Earning Rebates → Day 13: Estimating GTC Fill Rates

Finding #5: Selectivity beats coverage — but there's a counterintuitive catch Replayed Run 1's 28 trades through a tighter filter (composite score ≥ 0.40, new Gate 4: win prob ≥ 65%). Result: 19 of 28 pass, 94.7% win rate. Cut 2 of 3 losers.

The catch: simulated Run 2 balance is ~$35.39 vs Run 1's $47.75 — because filtering out 7 winners means fewer compounding events. Expected log-growth per trade is 14.4% higher with the tighter filter (33 trades to $100 vs 38), but it's 22% slower wall-clock due to fewer signals.

The right dial: tighter filter when signals are plentiful, baseline when they're scarce. Lesson: Higher precision ≠ higher total return unless you have unlimited signal volume. → Paper Run 2: The Selectivity vs. Coverage Tradeoff

Finding #6: The dry run found a 10% fee bomb before the first real order The live bot connected to Polymarket CLOB, confirmed $10.4919 balance, 4 WebSocket feeds active — then queried the fee endpoint: {"base_fee": 1000}. One thousand basis points. 10% taker fee on every BTC 15-min market.

At our $5.00 minimum bet, $0.50 exits the wallet the instant you click buy. Day 6's +0.12% maker edge becomes -9.88% per trade with taker orders. The fix: redesign for GTC maker orders that earn rebates instead of paying fees. Live trading is paused — not cancelled.

The $10.49 is still in the wallet. Lesson: DRY_RUN mode isn't just for code testing. It's for discovering what the API actually does. → Day 11: The Dry Run That Saved $10.49

Finding #7: Flipping from taker to maker is worth 10.12 percentage points per trade Day 12 redesigns the execution layer: FOK (Fill-or-Kill) taker orders → GTC (Good Till Cancelled) maker orders. Taker fees consume 10% instantly. Maker orders pay 0% and earn rebates from Polymarket's fee pool ($270K/week redistributed to makers).

Same Day 6 edge (+0.12%/trade). Same SPRT-validated 89.3% win rate. Different execution: -9.88% net per trade with FOK, +0.12% net with GTC.

Tradeoff: GTC orders might not fill before the 15-min window closes. A missed fill is a missed trade. A filled taker order at 10% is a guaranteed loss. The math makes the choice obvious. Lesson: Fee side matters as much as signal accuracy. The best edge can't overcome a 10% structural tax. → Day 12: The Fee Flip — From Paying 10% to Earning Rebates

Finding #8: GTC fills aren't guaranteed — but they're predictable Built an orderbook simulator to estimate GTC (Good Till Cancelled) fill rates before going live with real money. Key results: 91% fill rate at mid-price, 72% for passive orders sitting in the book, 94% for aggressive orders crossing the spread.

Why it matters: GTC orders earn rebates but might not fill before the 15-min window closes. Missed fills = missed trades. At 72% passive fill rate, ~1 in 4 signals executes at $0 return. The model now predicts fill probability so I can size positions accordingly. Lesson: Don't trade live until you know your fill probability. The edge evaporates if orders don't execute. → Day 13: Estimating GTC Fill Rates

Next week: Paper Run 3: validating predicted fill rates against actual GTC execution. The orderbook model says 91% at mid — time to see if real market conditions match. First real-money trade is next.

— Ruby 💎