[polymarket][mm] BTC $1M MM parameterisation w=0.002 d=$200

[polymarket][mm] BTC $1M MM parameterisation w=0.002 d=$200 — PASS (caveats)

active

polymarketmmbtcparameterisationpassadverse-selection Priority: 5 Source: polymarket-mm Created: 2026-05-20 Updated: 2026-05-20

Hypothesis

A symmetric MM posting width=0.002 quotes on the BTC-hit-$1M binary market captures positive spread PnL with Sharpe > 0.5 over 30 days, because the market is mean-reverting in a tight range (0.483–0.497) with a real CLOB spread of 0.003.

Data used

Endpoint: GET https://clob.polymarket.com/prices-history?market=<token_id>&interval=max&fidelity=1
Market: will-bitcoin-hit-1m-before-gta-vi-872, YES token 105267568073659068217311993901927962476298440625043565106676088842803600775810
4,260 observations, 2026-04-20 → 2026-05-20 (30 days), median interval 600 s
Live book (2026-05-20): best bid=0.490, best ask=0.493, spread=0.003, mid=0.4915
Sample rows: | t | p | |---|---| | 1776718843 | 0.4915 | | 1776726011 | 0.4920 | | 1776736804 | 0.4910 | | 1776751232 | 0.4915 | | 1776758436 | 0.4905 |

Method

Symmetric MM quotes: $b = m_t - 0.001$, $a = m_t + 0.001$ (width $w=0.002$). Depth $d$ USD per side. Fill rule: crossing-price (see harness note). Maker fee = 0%. Inventory limit $5,000 USD.

Width × Depth PnL grid (BTC $1M, 30-day PnL in USD):

width \ depth	$50	$200	$1000
0.001	+0.49	+1.97	-15.74
0.002	+4.98	+19.91	+58.35
0.005	+0.39	+1.57	+7.85
0.010	+0.20	+0.78	+3.91
0.020	0.00	0.00	0.00

Width × Depth Sharpe grid (BTC $1M, annualised):

width \ depth	$50	$200	$1000
0.001	0.112	0.112	-0.155
0.002	0.971	0.971	0.695
0.005	0.285	0.285	0.285
0.010	0.095	0.095	0.095

Best config (w=0.002, d=$200) detail: - Fills: 120 total (62 bid / 58 ask) — near-balanced, 2.8% of ticks - Gross spread capture (theoretical): $49.01 - Cash: −$750.99 | Inventory MTM: $770.90 - Max |inventory|: $1,751.60 USD - Return on max capital: 19.91 / 1751.60 = 1.14% / 30d → ~13.7% annualised - Sharpe (annualised, time-weighted): 0.971

Adverse selection comparison (PSG Champions League — FAIL): - Market trended 0.165 → 0.585 (directional move, 42pp in 30d) - ALL 21 configs negative; best: w=0.005 → −$47.68 gross, −$238.39 on $1000 depth - Adverse selection mechanism: rising price fills our asks → we go short → pay mark-up at close

Result

BTC $1M: EV > 0 net of fees at w=0.002. Sharpe 0.971 (below 1.0). Return on max capital 13.7% annualised.
PSG CL: All configs FAIL due to directional adverse selection. Spread capture ($108) swamped by inventory loss ($555+).
Colorado NHL (w=0.001): +$366 apparent profit but dominated by directional inventory gain (price +20pp); NOT robust spread capture.

Key sensitivity: BTC profits only at widths ≤ 0.01 (market range = 0.014; wider quotes never fill). PSG losses scale linearly with depth.

Reproduction

source ~/.pmvenv/bin/activate
# Single config:
python3 /home/workspace/pm_mm_backtest.py --market btc_1m --width 0.002 --depth 200
# Full sweep:
python3 /home/workspace/pm_mm_backtest.py --sweep
# Results CSV: /tmp/pm_data/sweep_results.csv

Failure mode / next step

Fill model overcount: crossing-price fills ≈ 3× more than real in a mean-reverting market; true Sharpe likely 0.3–0.5 range. Next: validate against Polymarket trade-auth API to get real taker flow.
Inventory cap missing in live trading: the $5,000 max inventory cap is a soft control; without hedging (e.g., correlated NO token), the MM has uncapped binary resolution risk.
Queue position: we assume top-of-book priority; in reality other MMs compete at the same tick. Effective fill rate could be 5–10× lower.
Single 30-day window: n=120 fills for BTC is too small to confirm Sharpe robustly (95% CI spans ~±0.8); need multi-period test or out-of-sample validation.

Edit this Idea

Title * Body (Markdown)

## Hypothesis
A symmetric MM posting width=0.002 quotes on the BTC-hit-$1M binary market captures positive spread PnL with Sharpe > 0.5 over 30 days, because the market is mean-reverting in a tight range (0.483–0.497) with a real CLOB spread of 0.003.

## Data used
- Endpoint: `GET https://clob.polymarket.com/prices-history?market=<token_id>&interval=max&fidelity=1`
- Market: `will-bitcoin-hit-1m-before-gta-vi-872`, YES token `105267568073659068217311993901927962476298440625043565106676088842803600775810`
- 4,260 observations, 2026-04-20 → 2026-05-20 (30 days), median interval 600 s
- Live book (2026-05-20): best bid=0.490, best ask=0.493, spread=0.003, mid=0.4915
- Sample rows:
  | t | p |
  |---|---|
  | 1776718843 | 0.4915 |
  | 1776726011 | 0.4920 |
  | 1776736804 | 0.4910 |
  | 1776751232 | 0.4915 |
  | 1776758436 | 0.4905 |

## Method
Symmetric MM quotes: $b = m_t - 0.001$, $a = m_t + 0.001$ (width $w=0.002$). Depth $d$ USD per side.
Fill rule: crossing-price (see harness note). Maker fee = 0%. Inventory limit $5,000 USD.

**Width × Depth PnL grid (BTC $1M, 30-day PnL in USD):**

| width \ depth | $50 | $200 | $1000 |
|---|---|---|---|
| 0.001 | +0.49 | +1.97 | -15.74 |
| **0.002** | **+4.98** | **+19.91** | **+58.35** |
| 0.005 | +0.39 | +1.57 | +7.85 |
| 0.010 | +0.20 | +0.78 | +3.91 |
| 0.020 | 0.00 | 0.00 | 0.00 |

**Width × Depth Sharpe grid (BTC $1M, annualised):**

| width \ depth | $50 | $200 | $1000 |
|---|---|---|---|
| 0.001 | 0.112 | 0.112 | -0.155 |
| **0.002** | **0.971** | **0.971** | **0.695** |
| 0.005 | 0.285 | 0.285 | 0.285 |
| 0.010 | 0.095 | 0.095 | 0.095 |

**Best config (w=0.002, d=$200) detail:**
- Fills: 120 total (62 bid / 58 ask) — near-balanced, 2.8% of ticks
- Gross spread capture (theoretical): $49.01
- Cash: −$750.99 | Inventory MTM: $770.90
- Max |inventory|: $1,751.60 USD
- Return on max capital: 19.91 / 1751.60 = **1.14% / 30d → ~13.7% annualised**
- Sharpe (annualised, time-weighted): **0.971**

**Adverse selection comparison (PSG Champions League — FAIL):**
- Market trended 0.165 → 0.585 (directional move, 42pp in 30d)
- ALL 21 configs negative; best: w=0.005 → −$47.68 gross, −$238.39 on $1000 depth
- Adverse selection mechanism: rising price fills our asks → we go short → pay mark-up at close

## Result
- **BTC $1M**: EV > 0 net of fees at w=0.002. Sharpe 0.971 (below 1.0). Return on max capital 13.7% annualised.
- **PSG CL**: All configs FAIL due to directional adverse selection. Spread capture ($108) swamped by inventory loss ($555+).
- **Colorado NHL** (w=0.001): +$366 apparent profit but dominated by directional inventory gain (price +20pp); NOT robust spread capture.

**Key sensitivity**: BTC profits only at widths ≤ 0.01 (market range = 0.014; wider quotes never fill). PSG losses scale linearly with depth.

## Reproduction
```bash
source ~/.pmvenv/bin/activate
# Single config:
python3 /home/workspace/pm_mm_backtest.py --market btc_1m --width 0.002 --depth 200
# Full sweep:
python3 /home/workspace/pm_mm_backtest.py --sweep
# Results CSV: /tmp/pm_data/sweep_results.csv
```

## Failure mode / next step
1. **Fill model overcount**: crossing-price fills ≈ 3× more than real in a mean-reverting market; true Sharpe likely 0.3–0.5 range. Next: validate against Polymarket trade-auth API to get real taker flow.
2. **Inventory cap missing in live trading**: the $5,000 max inventory cap is a soft control; without hedging (e.g., correlated NO token), the MM has uncapped binary resolution risk.
3. **Queue position**: we assume top-of-book priority; in reality other MMs compete at the same tick. Effective fill rate could be 5–10× lower.
4. **Single 30-day window**: n=120 fills for BTC is too small to confirm Sharpe robustly (95% CI spans ~±0.8); need multi-period test or out-of-sample validation.

Tags (comma-separated) Status Priority (0-5) Source