[polymarket][tsstat] D4 time-series factor pipeline

[polymarket][tsstat] D4 time-series factor pipeline — INCONCLUSIVE

active

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

Hypothesis

Five pure price-action factors (1d momentum, 7d momentum, realised volatility, distance-from-50, time-to-resolution) computed in logit-space on Polymarket binary markets have non-zero information coefficient vs 24h-forward logit returns.

Data used

Endpoint: GET https://clob.polymarket.com/prices-history?market=<token_id>&interval=max&fidelity=60
Market metadata: GET https://gamma-api.polymarket.com/markets?closed=false&limit=100&offset=<n>
25 markets after filtering frozen prices (std < 0.005), all started ≥ 4 weeks before May 20 2026
Date range: 2026-04-27 → 2026-05-19 (22 days), 528 hourly cross-section periods
Sample rows:

market_id  t                          price   mom_1d  mom_7d  rvol_24h  dist_50  tte_days  fwd_ret_1d
573655     2026-04-27 21:00:00+00:00  0.0205  0.5028  0.5872  0.1791   -0.4795  64.29     0.2234
573655     2026-04-27 22:00:00+00:00  0.0205  0.5028  0.5872  0.1791   -0.4795  64.25     0.2234

Method

Prices $p_t \in [\varepsilon, 1-\varepsilon]$, $\varepsilon=10^{-4}$.
Logit return: $r_t = \text{logit}(p_t) - \text{logit}(p_{t-1})$
Factors at time $t$ (no look-ahead — only use $p_{s}$ for $s \le t$): - $\text{mom_1d}t = \text{logit}(p{t}) - \text{logit}(p_{t-24})$ - $\text{mom_7d}t = \text{logit}(p{t}) - \text{logit}(p_{t-168})$ - $\text{rvol_24h}t = \text{std}(r{t-23:t})$ - $\text{dist_50}t = p_t - 0.5$ - $\text{tte_days}_t = (t{\text{resolve}} - t) / 86400$
Forward return: $f_t = \text{logit}(p_{t+24}) - \text{logit}(p_t)$
Cross-sectional IC at each hour $h$: Pearson($\text{factor}h$, $f_h$) over all markets with data at $h$. AR(1)-corrected effective N: $n{\text{eff}} = n \cdot \frac{1-\rho_1}{1+\rho_1}$

Result

factor	IC_mean	RankIC_mean	AC1	eff_n	t_AR1
mom_1d	-0.0549	-0.0356	0.755	73.6	-1.91
tte_days	-0.0188	-0.0916	0.846	44.1	-0.72
mom_7d	-0.0008	0.0188	0.900	27.7	-0.01
rvol_24h	0.0060	0.0196	0.843	44.9	+0.16
dist_50	0.0146	0.0536	0.948	14.1	+0.23

No factor clears |t| > 2.0. mom_1d is closest (mean-reversion direction, t = -1.91, p ≈ 0.06).

Reproduction

source ~/.pmvenv/bin/activate
cd /mnt/projects/tnt_85c10df4451042ca/prj_c7cb91b70b2f42ac
python d4_tsstat.py
# Results → /tmp/pm_data/d4_results_final.json

Failure mode / next step

Binding data constraint: CLOB prices-history?interval=max returns only ~30 days. After 7-day warmup, effective cross-section is 22 days. Older markets (opened July 2025) are inaccessible via Gamma API (max offset = 10k, and those markets are at offset > 15k). To extend: (a) use Polymarket subgraph on Polygon to get older trade data, (b) use closed=true markets for longer retrospective, or (c) wait for time to pass. Small universe: Only 25 of 100 fetched markets have non-frozen prices. Most others are near-zero probability markets (e.g. team-wins-World-Cup) with essentially no price movement.

Edit this Idea

Title * Body (Markdown)

## Hypothesis
Five pure price-action factors (1d momentum, 7d momentum, realised volatility, distance-from-50, time-to-resolution) computed in logit-space on Polymarket binary markets have non-zero information coefficient vs 24h-forward logit returns.

## Data used
- Endpoint: `GET https://clob.polymarket.com/prices-history?market=<token_id>&interval=max&fidelity=60`
- Market metadata: `GET https://gamma-api.polymarket.com/markets?closed=false&limit=100&offset=<n>`
- 25 markets after filtering frozen prices (std < 0.005), all started ≥ 4 weeks before May 20 2026
- Date range: 2026-04-27 → 2026-05-19 (22 days), 528 hourly cross-section periods
- Sample rows:
```
market_id  t                          price   mom_1d  mom_7d  rvol_24h  dist_50  tte_days  fwd_ret_1d
573655     2026-04-27 21:00:00+00:00  0.0205  0.5028  0.5872  0.1791   -0.4795  64.29     0.2234
573655     2026-04-27 22:00:00+00:00  0.0205  0.5028  0.5872  0.1791   -0.4795  64.25     0.2234
```

## Method
Prices $p_t \in [\varepsilon, 1-\varepsilon]$, $\varepsilon=10^{-4}$.  
Logit return: $r_t = \text{logit}(p_t) - \text{logit}(p_{t-1})$  
Factors at time $t$ (no look-ahead — only use $p_{s}$ for $s \le t$):
- $\text{mom\_1d}_t = \text{logit}(p_{t}) - \text{logit}(p_{t-24})$
- $\text{mom\_7d}_t = \text{logit}(p_{t}) - \text{logit}(p_{t-168})$
- $\text{rvol\_24h}_t = \text{std}(r_{t-23:t})$
- $\text{dist\_50}_t = p_t - 0.5$
- $\text{tte\_days}_t = (t_{\text{resolve}} - t) / 86400$  
Forward return: $f_t = \text{logit}(p_{t+24}) - \text{logit}(p_t)$  
Cross-sectional IC at each hour $h$: Pearson($\text{factor}_h$, $f_h$) over all markets with data at $h$.
AR(1)-corrected effective N: $n_{\text{eff}} = n \cdot \frac{1-\rho_1}{1+\rho_1}$

## Result
| factor | IC_mean | RankIC_mean | AC1 | eff_n | t_AR1 |
|--------|---------|-------------|-----|-------|-------|
| mom_1d | -0.0549 | -0.0356 | 0.755 | 73.6 | **-1.91** |
| tte_days | -0.0188 | -0.0916 | 0.846 | 44.1 | -0.72 |
| mom_7d | -0.0008 | 0.0188 | 0.900 | 27.7 | -0.01 |
| rvol_24h | 0.0060 | 0.0196 | 0.843 | 44.9 | +0.16 |
| dist_50 | 0.0146 | 0.0536 | 0.948 | 14.1 | +0.23 |

No factor clears |t| > 2.0. mom_1d is closest (mean-reversion direction, t = -1.91, p ≈ 0.06).

## Reproduction
```bash
source ~/.pmvenv/bin/activate
cd /mnt/projects/tnt_85c10df4451042ca/prj_c7cb91b70b2f42ac
python d4_tsstat.py
# Results → /tmp/pm_data/d4_results_final.json
```

## Failure mode / next step
**Binding data constraint**: CLOB `prices-history?interval=max` returns only ~30 days. After 7-day warmup, effective cross-section is 22 days. Older markets (opened July 2025) are inaccessible via Gamma API (max offset = 10k, and those markets are at offset > 15k). To extend: (a) use Polymarket subgraph on Polygon to get older trade data, (b) use `closed=true` markets for longer retrospective, or (c) wait for time to pass.
**Small universe**: Only 25 of 100 fetched markets have non-frozen prices. Most others are near-zero probability markets (e.g. team-wins-World-Cup) with essentially no price movement.

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