Polymarket Lab Dashboard — Arthur Breguez

polymarket-lab — bash — 220×48

██████╗  ██████╗ ██╗     ██╗   ██╗███╗   ███╗ █████╗ ██████╗ ██╗  ██╗███████╗████████╗    ██╗      █████╗ ██████╗
██╔══██╗██╔═══██╗██║     ╚██╗ ██╔╝████╗ ████║██╔══██╗██╔══██╗██║ ██╔╝██╔════╝╚══██╔══╝    ██║     ██╔══██╗██╔══██╗
██████╔╝██║   ██║██║      ╚████╔╝ ██╔████╔██║███████║██████╔╝█████╔╝ █████╗     ██║       ██║     ███████║██████╔╝
██╔═══╝ ██║   ██║██║       ╚██╔╝  ██║╚██╔╝██║██╔══██║██╔══██╗██╔═██╗ ██╔══╝     ██║       ██║     ██╔══██║██╔══██╗
██║     ╚██████╔╝███████╗   ██║   ██║ ╚═╝ ██║██║  ██║██║  ██║██║  ██╗███████╗   ██║       ███████╗██║  ██║██████╔╝
╚═╝      ╚═════╝ ╚══════╝   ╚═╝   ╚═╝     ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝╚══════╝   ╚═╝       ╚══════╝╚═╝  ╚═╝╚═════╝

artbreguez@lab:~/polymarket-lab$ status --all

● F1 champion MultitaskQR r4 PnL=+$21.91 ROI=64.3% bets=144 [ACTIVE]

● TMAX champion lgbm_emos gate=GO qp_pnl=+$138 [ACTIVE]

● GFS backfill 2602/2602 markets +1350 new forecasts 30 cities [DONE ✓]

● ECMWF backfill queued → triggers full retrain pipeline [PENDING]

● Miami 2026 Sprint weekend FP1: 2026-05-01 16:00 UTC [T-MINUS]

F1 Walk-Forward PnL

+$21.91

27 GPs · 144 bets

F1 ROI

64.3%

multitask_qr r4

TMAX Gate PnL

+$201

Atlanta + BA + Madrid

TMAX qp_pnl

+$138

live sim · lgbm_emos

Autoresearch Iters

450

3 seeds · r4 held

Training Cities

30

GFS: 2602 markets

// miami 2026 grand prix — countdown

SPRINT WEEKEND

FP1 start → 2026-05-01 16:00 UTC

--

days

:

--

hrs

:

--

min

:

--

sec

● champion r4 ready 5 crons scheduled paper trading only

// process status

GFS Backfill — 30 cities

2602/2602 · 1350 new · 0 errors

DONE

ECMWF Backfill (ifs025 + aifs025)

queued → full retrain pipeline

QUEUED

F1 Autoresearch v3 (3×150 iters)

450 candidates · r4 champion held

DONE

TMAX Champion Publish + HF Upload

lgbm_emos → huggingface/artbreguez

DONE

Miami 2026 Paper Trade Crons

FP1→SQ · SQ→Sprint · Q→Race · settlements

SCHEDULED

Post-GP Retrain Cron (Sundays)

auto sync + retrain + HF publish

ACTIVE

// F1 Lab / Champion Model

🔒

This section is private.
Click the lock icon in the top bar to authenticate.

multitask_qr r4

run_id: c21e58b0-29c4-544a-97a7-f7ac9b763ee3 · promoted: 2026-04-28 12:26 UTC

+$21.91

Walk-Forward PnL

64.3%

ROI

144

Total Bets

hidden64

depth1

dropout0.05

lr5e-3

h2h_wt1.6

winner_wt0.4

pole_wt1.2

cpole_wt1.4

batch64

// per-family performance

Market Family	PnL	ROI	Bets	Note
head_to_head (H2H)	+$14.13	47%	60	primary alpha source
race_winner	+$5.62	166%	56	winner_wt=0.4 wins
driver_pole	+$2.16	257%	28	high precision

// challengers — miami 2026

autoresearch v3

Rank	PnL	ROI	Bets	hidden	depth	lr	h2h	winner	seed
#1	+$18.05	42%	175	64	2	0.005	1.6	0.4	7
#2	+$17.63	48%	130	80	1	0.005	1.4	0.4	99
#3	+$16.23	48%	149	160	3	0.002	1.8	0.8	7

// key insights

▸

H2H is the alpha source. h2h_weight=1.6 in every top config. Polymarket H2H prices are systematically less efficient than outright markets. 73% of total PnL comes from H2H bets.

▸

Shallow beats deep. depth=1 outperforms depth=3-4. Only 27 GPs of data — large models overfit. Champion is the simplest architecture that survived 450 iterations.

▸

winner_weight=0.4 wins. Deliberately downweighting winner head forces the shared encoder to learn representations useful for H2H. Counterintuitive but confirmed across 7 rounds.

▸

Dataset ceiling reached. Polymarket had no F1 markets before 2024 R10. 94 snapshots / 27 GPs is the maximum possible. More autoresearch won't find dramatically better models without more data.

// huggingface backup

SYNCED

Artifacts auto-uploaded after every champion promotion.

repo: artbreguez/f1-polymarket-champion (private)
files: model_bundle.pt · calibrators.json · trainer_config.json · metrics.json
trigger: post_gp_retrain.sh → publish_champion_to_hf.py --project f1

⬡ huggingface.co/artbreguez/f1-polymarket-champion ↗

// F1 Lab / Autoresearch History

🔒

This section is private.
Click the lock icon in the top bar to authenticate.

// autoresearch progression — 7 rounds

Round	Iters	Method	Best PnL	vs r4	Result
r1	20	random	+$15.97	-27%	baseline found
r2	30	random expanded	+$15.94	-27%	plateau
r3	40	hill-climb intro	+$18.78	-14%	NEW RECORD
r4	40	hill-climb seeded	+$21.91	—	★ CHAMPION
r5	40	grids too wide	+$17.07	-22%	regression
r6	50	ultra-fine	+$19.73	-10%	convergence confirmed
r7	450	3× parallel seeds	+$18.05	-18%	r4 held · ceiling reached

// hill-climbing algorithm (r3+)

# 70% neighbourhood search + 30% random jump
for i in range(iterations):
    if rng.random() < 0.70:
        candidate = hill_climb_config(rng, best_seed)  # ±1 grid step
    else:
        candidate = mutate_config(rng, baseline)       # random jump
    pnl = evaluate(candidate)
    if pnl > best_pnl:
        best_pnl = pnl
        best_seed = candidate  # update seed dynamically

After r4: grids tightened to ±2 steps around champion config. Confirmed convergence — next improvement requires new feature engineering, not more search.

// F1 Lab / Miami 2026 — Sprint Weekend

// session schedule

SPRINT WEEKEND

01/May 16:00 UTC

FP1 — Free Practice 1

Cron fires 16:30 UTC → build snapshot → score → paper trade

SQ pole markets + Q pole markets

01/May 20:30 UTC

SQ — Sprint Qualifying

Cron fires 21:00 UTC → Sprint winner paper trade

miami_sq_sprint

02/May 16:00 UTC

Sprint Race

Settle miami_fp1_sq + miami_sq_sprint

settle

02/May 20:00 UTC

Q — Qualifying

Cron fires 22:30 UTC → settle Q pole

miami_fp1_q

03/May 20:00 UTC

Race

Q→Race paper trade + settle miami_q_r 2h after

miami_q_r

// paper trade slugs

Slug	Trigger	Target Session	Settle after
miami_fp1_sq	FP1 → 16:30	SQ pole	Sprint
miami_sq_sprint	SQ → 21:00	Sprint winner	Sprint result
miami_fp1_q	FP1 → 16:30	Q pole	Qualifying
miami_q_r	Q → 22:30	Race winner	Race result

// bankroll config

# paper trading parameters
bet_size    = $5.0      # min $5 to avoid CLOB rejection (<$1 orders fail)
min_edge    = 0.05      # 5% edge minimum
order_type  = 'GTC'     # Good Till Cancelled
meeting_key = 1284      # Miami 2026 (positive key = 2026+)
season      = 2026

// F1 Lab / PnL History — Walk-Forward

// cumulative pnl by gp (r4 champion)

27 GPs

Unitized PnL (1 unit = 1 share). Multiply × bet_size for dollars.

2024 R122025 R12025 R6 →

// per-family cumulative pnl

Family	Total PnL	ROI	Bets	Hit Rate
head_to_head	+$14.13	47%	60	—
race_winner	+$5.62	166%	56	—
driver_pole	+$2.16	257%	28	—
TOTAL	+$21.91	64.3%	144	—

// TMAX Lab / Champion Model

lgbm_emos [alias: champion]

promoted: 2026-04-29 14:51 UTC · gate: GO · 3 cities passed

+$138

qp_pnl (live sim)

+$201

gate aggregate

0.1137

Brier Score

// model leaderboard

historical backtest

Model	PnL	qp_pnl	Brier	Trades	Status
lgbm_emos (tuned_ensemble)	+$381	+$138	0.1137	1245	CHAMPION
gaussian_emos	+$461	+$4	0.1119	1242	NO EDGE
det2prob_nn	+$25	-$173	0.1176	1254	SKIP

// recent-core gate — city results

GO

City	Trades	PnL	ok_ratio	Gate
Atlanta	76	+$162.45	100%	PASS ✓
Buenos Aires	76	+$28.84	100%	PASS ✓
Madrid	66	+$10.25	100%	PASS ✓

// key metric: qp_pnl > brier

!

gaussian_emos has better Brier (0.1119) but qp_pnl ≈ $4. A well-calibrated model that matches Polymarket prices offers no tradeable edge. Always select champion by quote_proxy_pnl, not raw Brier score.

▸

London/NYC failed the gate (negative PnL, z-scores -5.57 and -8.08). Not noise. Switched to Atlanta/Buenos Aires/Madrid for recent-core validation.

// TMAX Lab / Retrain Pipeline

🔒

This section is private.
Click the lock icon in the top bar to authenticate.

// full retrain pipeline — current run

IN PROGRESS

GFS Backfill — all 30 cities

2602/2602 · 1350 new rows · 0 errors

DONE

ECMWF Backfill (ifs025 + aifs025)

adds neighbor_spread feature to all rows

expected: ~8,000+ rows (was 6,447)

PENDING

Re-materialize backtest panel

30 cities · ok-only filter

PENDING

Retrain all models

gaussian_emos · 4× lgbm_emos · tuned_ensemble

PENDING

Benchmark (stride=30)

leaderboard.json → qp_pnl ranking

PENDING

Recent-core gate (stride=1)

Atlanta · Buenos Aires · Madrid

PENDING

Publish + HF upload

if GO → publish-champion → huggingface

PENDING

// training set history

Before (GFS only, 20 cities)	6,447 rows
After GFS backfill (30 cities)	~7,800 rows (est.)
After ECMWF backfill (neighbor_spread)	~7,800 rows + feature

neighbor_spread = |ecmwf_ifs025 − ecmwf_aifs025_single| → uncertainty signal. Was 0 for all rows before this run.

// Ops / VPS Recovery Guide

🔒

This section contains private infrastructure details.
Click the lock icon in the top bar to authenticate.

!! READ FIRST — model backups on HuggingFace

Champion models are backed up automatically after every promotion. Download before rebuilding:

# F1 champion
pip install huggingface_hub
python3 -c "from huggingface_hub import snapshot_download; snapshot_download('artbreguez/f1-polymarket-champion', token='HF_TOKEN', local_dir='/tmp/f1')"

# TMAX champion
python3 -c "from huggingface_hub import snapshot_download; snapshot_download('artbreguez/tmax-polymarket-champion', token='HF_TOKEN', local_dir='/tmp/tmax')"

1

System dependencies

sudo apt-get install -y postgresql redis-server git curl build-essential curl -LsSf https://astral.sh/uv/install.sh | sh

2

Start services + create DB

sudo systemctl start postgresql redis sudo -u postgres psql -c "ALTER USER postgres PASSWORD 'PASS';" sudo -u postgres psql -c "CREATE DATABASE f1_polymarket_lab OWNER postgres;"

3

Clone repos

git clone <f1-repo-url> ~/f1-polymarket-lab git clone <tmax-repo-url> ~/polymarket-tmax-lab git clone https://github.com/ArtBreguez/f1-weather-lab-wiki.git ~/f1-weather-lab-wiki

4

Configure F1 lab (.env)

cp ~/f1-polymarket-lab/.env.example ~/f1-polymarket-lab/.env

Edit: POSTGRES_PASSWORD, DATA_ROOT, MLFLOW_TRACKING_URI=http://127.0.0.1:5001

5

Install deps + migrate DB

cd ~/f1-polymarket-lab && uv sync --all-packages --group dev --group modeling && make db-upgrade cd ~/polymarket-tmax-lab && uv sync --all-extras

6

Start MLflow

cd ~/f1-polymarket-lab && mkdir -p data/mlflow/artifacts && uv run mlflow server --backend-store-uri "sqlite:///$(pwd)/data/mlflow/mlflow.db" --default-artifact-root "$(pwd)/data/mlflow/artifacts" --host 127.0.0.1 --port 5001 &

7

Rebuild datasets (takes hours — use --forecast-missing-only to resume)

cd ~/polymarket-tmax-lab && scripts/pmtmax-workspace historical_real uv run pmtmax build-dataset --markets-path configs/market_inventory/full_training_set_snapshots.json --forecast-missing-only --allow-canonical-overwrite > /tmp/tmax_rebuild.log 2>&1 &

8

Recreate crons via Hermes agent

Ask Hermes to recreate: f1-post-gp-retrain (Sun 22h), tmax-weekly-retrain (Wed 6h), disk-monitor (6h), wiki-auto-update (6h). See cron registry page for exact prompts.

9

Health check

pg_isready -h 127.0.0.1 && redis-cli ping && curl -s http://127.0.0.1:5001/health cd ~/f1-polymarket-lab && python scripts/health_check_miami.py

// Ops / Cron Registry

🔒

This section contains private cron job IDs and infrastructure details.
Click the lock icon in the top bar to authenticate.

// active cron jobs

5 ACTIVE

Name	Job ID	Schedule	Status
f1-post-gp-retrain	e9541147b363	Sun 22:00 UTC	ACTIVE
tmax-weekly-retrain	a3e98dc8d385	Wed 06:00 UTC	ACTIVE
tmax-full-retrain-pipeline	c8d995a44e66	every 20min (12×)	TEMP
disk-monitor	disk-monitor	every 6h	ACTIVE
wiki-auto-update	wiki-update	every 6h	ACTIVE

// what each cron does

f1-post-gp-retrain — Sundays 22:00 UTC

Detects new completed F1 races → Jolpica sync → catalog sync → reconcile mappings → build snapshots → rebuild manifest → fix schemas → walk-forward retrain (r4 config) → aggregate folds → promote if gate pass → HuggingFace upload → wiki push

tmax-weekly-retrain — Wednesdays 06:00 UTC

backfill-truth (new resolved markets) → build-dataset (new forecasts) → re-materialize training set → if +100 new rows: full retrain pipeline (ECMWF + all models + benchmark + gate + promote + HF upload)

disk-monitor — every 6h

Warn at 70% disk. Auto-clean safe targets (TMAX cache, /tmp logs) at 85%.

// log locations

# F1 post-GP retrain
/tmp/f1_post_gp_retrain/retrain_*.log

# TMAX full pipeline
/tmp/tmax_pipeline/full_pipeline.log
/tmp/tmax_pipeline/ecmwf_backfill.log
/tmp/tmax_pipeline/recent_core.log
/tmp/tmax_pipeline/hf_upload.log

# GFS backfill
/tmp/tmax_gfs_backfill.log