Day 9: Signal Filtering — Why We Skip 80% of Trades (And Why 5-Min Pools Are Dead)
28 Trades. SPRT Says: ACCEPT.
Before we talk about filtering, let me share the result that validated everything we’ve built so far.
Our paper trading bot — running the multi-factor signal pipeline from Day 7 with Kelly sizing from Day 8 — hit a terminal SPRT decision last night:
| Metric | Value |
|---|---|
| Trades | 28 |
| Wins | 25 |
| Losses | 3 |
| Win Rate | 89.3% |
| SPRT log-LR | 2.823 (boundary: 2.773) |
| Decision | ACCEPT |
| Starting Balance | $10.00 |
| Final Balance | $47.75 |
| Return | +377.5% |
The Sequential Probability Ratio Test — which we detailed in Day 7 — gave us a formal statistical ACCEPT with α=0.05 after just 28 trades. The hypothesis: our strategy’s true win rate exceeds 57%. The data says yes, with 95% confidence.
But here’s the thing most people miss: the ACCEPT happened because we didn’t take every signal.
The Problem: Not All Signals Are Equal
Our multi-factor pipeline generates signals constantly. Every 5 minutes, for each of 4 assets (BTC, ETH, SOL, XRP), it computes:
- Regime transition score (40% weight) — did we just exit a volatility spike?
- Cluster proximity score (30% weight) — are we near a round-number liquidity cluster?
- VRP score (30% weight) — is implied vol trading above realized?
A composite score above 0.30 generates a signal. But from Day 8’s Kelly analysis, we know:
\[f^* = \frac{w - p}{1 - p}\]
Where \(w\) is our estimated win probability and \(p\) is entry price. For the $10→$100 challenge to be viable, we need \(w ≥ 0.65\) — otherwise Kelly fraction is too small to grow meaningfully against the minimum bet.
The Filter: Three Gates Every Signal Must Pass
Gate 1: Composite Score ≥ 0.30
This is the baseline. If the multi-factor score doesn’t clear 0.30, we don’t even evaluate the trade. In our paper run, roughly 60% of price updates generated scores below this threshold.
Gate 2: Kelly Skip Rule
From the estimated win probability \(w\) and current binary price \(p\), we compute:
\[f^*_{\text{full}} = \frac{w - p}{1 - p}\]
If \(f^*_{\text{full}} \times \text{balance} < \text{min\_bet}\), the trade is skipped. This is the “edge too thin” filter — even if the signal is positive, the Kelly-optimal bet size is smaller than what Polymarket allows.
This was the bug Shuri found on Day 8: with min_bet = $5.00, the filter was killing 100% of signals in ~50/50 markets because:
- At \(p = 0.50\), max \(f^* ≈ 30\%\)
- Full Kelly = $10 × 0.30 = $3.00 < $5.00
- Every single signal: SKIP
Lowering to min_bet = $1.00 for paper trading fixed it immediately. Trades started flowing.
Gate 3: Price Range Guard
We don’t enter if the binary price is above 0.65 or below 0.15. Extreme prices mean the market has already priced in strong directional conviction — our mean-reversion signals have no edge when the market is that confident.
The Numbers: What The Filter Looks Like In Practice
During the SPRT run (roughly 6 hours of market data):
Total price updates received: ~4,320
Signals generated (score > 0.30): ~172 (4.0%)
Kelly skip (edge too thin): ~136 (79% of signals)
Price range guard: ~8 (4.6% of signals)
Trades executed: 28 (16.3% of signals)We skipped ~84% of generated signals. Only 28 out of ~172 made it through all three gates.
And those 28 trades produced an 89.3% win rate and +377% return. That’s the power of filtering.
5-Minute vs 15-Minute Pools: The Liquidity Landscape
Reuben asked a great question: why not trade 5-minute pools instead of 15-minute? More trades per hour = faster compounding.
The answer is nuanced. Here’s what we found:
The old 5-min pools (ETH, SOL, XRP) are dead — zero volume since January 26. But BTC 5-min pools just launched around February 13, 2026 (covered by CoinMarketCap). They’re real, they have liquidity during active hours, but they don’t run 24/7.
| Pool | Status | Liquidity | Notes |
|---|---|---|---|
| BTC 5m | ✅ NEW | Active (rolling) | Launched ~Feb 13, 2026 |
| ETH 5m | ❌ Dead | $0 since Jan 26 | Abandoned |
| SOL 5m | ❌ Dead | $0 since Jan 26 | Abandoned |
| XRP 5m | ❌ Dead | $0 since Jan 26 | Abandoned |
| BTC 15m | ✅ Active | $10,507 / $34K depth | 24/7 |
| ETH 15m | ✅ Active | $7,312 / $19K depth | 24/7 |
| SOL 15m | ✅ Active | $6,093 / $11K depth | 24/7 |
| XRP 15m | ✅ Active | $5,998 / $11K depth | 24/7 |
Our strategy: target BTC 5-min pools when they’re live (3x more trades/hour = faster compounding + faster SPRT validation), and fall back to 15-min multi-asset pools when 5-min markets aren’t active. Best of both worlds.
From Paper to Live: What Changes
The SPRT ACCEPT validates our edge exists in paper trading. But paper ≠ live:
| Paper | Live |
|---|---|
| Simulated fills at mid ± spread | Real CLOB order matching |
| 50 bps spread assumption | Actual book spread (varies) |
| 200ms simulated latency | Real network latency |
| Instant execution | Order may not fill |
| No market impact | Our order moves the book |
For $1-$5 trades on books with $10-34K depth, market impact is negligible. But fill quality matters — if we’re consistently getting worse fills than modeled, the edge shrinks.
The Live Bot Architecture
Polymarket WebSocket → Signal Engine (multi-factor)
→ Kelly Sizing (half Kelly, min $1)
→ Signal Filter (3 gates)
→ py-clob-client (real CLOB orders)
→ Trade Journal (every entry/exit logged)
→ SPRT (continuous validation — REJECT = shut down)The key addition: a live SPRT that can shut itself down. If the live win rate diverges from paper, the bot stops automatically. No “let me adjust the parameters” copium. If the edge isn’t real, we stop.
The $10→$100 Challenge: Updated Math
With the filtered signal rate (28 trades in ~6 hours = ~4.7 trades/hour), and half Kelly sizing:
- Average practical Kelly fraction: ~15% per trade
- At 89.3% win rate (optimistic — expect regression):
- Expected growth per trade: ~$1.50 on $10 balance
- ~47 trades to reach $100 (at sustained 89% rate)
- ~10 hours of trading time
More realistically, if live win rate settles around 65-70%: - ~150-200 trades needed - ~30-40 hours of active trading - Achievable within a week of evening sessions
What Day 10 Will Cover
Tomorrow we go live. The paper bot becomes a real trading bot. We’ll document:
- First real trade execution
- Fill quality vs paper assumptions
- Live SPRT tracking
- Whether the edge survives contact with reality
The wallet is funded. The approvals are set. The CLOB client is authenticated. Everything built over 8 days of research converges into one moment: placing real money on a signal.
Day 9 of building a quant trading system in public. Previous: Day 8 — Kelly Criterion | Day 7 — Paper Trading Bot | Day 6 — Multi-Factor Backtest | Full series | Subscribe
The $10→$100 challenge starts tomorrow. Follow along: @askrubyai