[polymarket][tsstat] realised volatility cross-sectional rank

[polymarket][tsstat] realised volatility cross-sectional rank — FAIL

active

polymarkettsstatvolatilityrvolfail Priority: 3 Source: polymarket-tsstat Created: 2026-05-20 Updated: 2026-05-20

Hypothesis

Markets with high realised volatility over the past 24h will continue to have higher expected absolute returns (volatility clustering), so going long high-rvol markets and short low-rvol markets captures a volatility risk premium or persistence effect.

Data used

Same 25 active Polymarket binary markets as [[D4-pipeline]]
Date range: 2026-04-27 → 2026-05-19, 528 hourly cross-section periods
Endpoint: https://clob.polymarket.com/prices-history?market=<token_id>&interval=max&fidelity=60

Method

$$\text{rvol_24h}t = \text{std}\left({r{t-23}, r_{t-22}, \ldots, r_t}\right), \quad r_s = \text{logit}(p_s) - \text{logit}(p_{s-1})$$ Long top quintile (highest rvol), short bottom quintile (lowest rvol) at each hour.
Transaction cost: $2\text{c}$ half-spread per new position leg, approximated in logit-space as $0.02 / (p(1-p)) \approx 0.08$ at $p=0.5$.
Net PnL = gross logit L/S spread minus TC on turnover.

Result

Metric	Value
Cross-sec IC mean	+0.006
Rank-IC mean	+0.020
AC(1) of IC series	0.843
Effective N (AR1)	44.9
t_AR1	+0.16
Gross L/S PnL (22d)	+1.64 logit units
Avg turnover	5.8% / period
TC drag	−5.21 logit units
Net PnL	−3.57 logit units
Ann. Sharpe (net)	−4.35
Hit rate (net)	0.449

Verdict: FAIL. IC is positive but not significant (t = 0.16). Gross PnL is the only bright spot: +1.64 logit units over 22 days with low turnover (5.8%). However, transaction costs of 2c/leg eat the entire gross profit, leaving net = -3.57.

The rvol persistence idea is plausible (volatility clustering is real in other asset classes), but the signal:cost ratio is too low for hourly rebalancing.

Reproduction

source ~/.pmvenv/bin/activate
python /mnt/projects/tnt_85c10df4451042ca/prj_c7cb91b70b2f42ac/d4_tsstat.py
# See: /tmp/pm_data/d4_results_final.json, factor=rvol_24h

Failure mode / next step

2c spread is too high for this signal size. At daily rebalance (not hourly), TC drops by ~24x. If gross PnL scales with fewer trades but same signal quality, daily rebalance might break even.
Signal direction is ambiguous: long high-rvol bets on continuation of active trading. But high-rvol markets may also be noisy rather than informative. Need to distinguish between rvol from price trend (directional) vs rvol from noise.
Cross-sectional IC AC1 = 0.843: The rvol ranking is very persistent — same markets lead volatility day after day. This is selection-bias-like: a few markets with structural uncertainty (elections, news events) dominate the top quintile always.
Next: Decompose rvol into trend-driven vs noise-driven components using realized skewness or autocorrelation of intraday returns.

Edit this Idea

Title * Body (Markdown)

## Hypothesis
Markets with high realised volatility over the past 24h will continue to have higher expected absolute returns (volatility clustering), so going long high-rvol markets and short low-rvol markets captures a volatility risk premium or persistence effect.

## Data used
- Same 25 active Polymarket binary markets as [[D4-pipeline]]
- Date range: 2026-04-27 → 2026-05-19, 528 hourly cross-section periods
- Endpoint: `https://clob.polymarket.com/prices-history?market=<token_id>&interval=max&fidelity=60`

## Method
$$\text{rvol\_24h}_t = \text{std}\left(\{r_{t-23}, r_{t-22}, \ldots, r_t\}\right), \quad r_s = \text{logit}(p_s) - \text{logit}(p_{s-1})$$
Long top quintile (highest rvol), short bottom quintile (lowest rvol) at each hour.  
Transaction cost: $2\text{c}$ half-spread per new position leg, approximated in logit-space as $0.02 / (p(1-p)) \approx 0.08$ at $p=0.5$.  
Net PnL = gross logit L/S spread minus TC on turnover.

## Result
| Metric | Value |
|--------|-------|
| Cross-sec IC mean | +0.006 |
| Rank-IC mean | +0.020 |
| AC(1) of IC series | 0.843 |
| Effective N (AR1) | 44.9 |
| **t_AR1** | **+0.16** |
| Gross L/S PnL (22d) | **+1.64 logit units** |
| Avg turnover | 5.8% / period |
| TC drag | −5.21 logit units |
| Net PnL | **−3.57 logit units** |
| Ann. Sharpe (net) | −4.35 |
| Hit rate (net) | 0.449 |

**Verdict: FAIL.** IC is positive but not significant (t = 0.16). Gross PnL is the only bright spot: +1.64 logit units over 22 days with low turnover (5.8%). However, transaction costs of 2c/leg eat the entire gross profit, leaving net = -3.57.

The rvol persistence idea is plausible (volatility clustering is real in other asset classes), but the signal:cost ratio is too low for hourly rebalancing.

## Reproduction
```bash
source ~/.pmvenv/bin/activate
python /mnt/projects/tnt_85c10df4451042ca/prj_c7cb91b70b2f42ac/d4_tsstat.py
# See: /tmp/pm_data/d4_results_final.json, factor=rvol_24h
```

## Failure mode / next step
1. **2c spread is too high for this signal size.** At daily rebalance (not hourly), TC drops by ~24x. If gross PnL scales with fewer trades but same signal quality, daily rebalance might break even.
2. **Signal direction is ambiguous**: long high-rvol bets on continuation of active trading. But high-rvol markets may also be noisy rather than informative. Need to distinguish between rvol from price trend (directional) vs rvol from noise.
3. **Cross-sectional IC AC1 = 0.843**: The rvol ranking is very persistent — same markets lead volatility day after day. This is selection-bias-like: a few markets with structural uncertainty (elections, news events) dominate the top quintile always.
4. **Next**: Decompose rvol into trend-driven vs noise-driven components using realized skewness or autocorrelation of intraday returns.

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