Kaggle Competition Type Strategy

The 6 Competition Types

Kaggle competitions are NOT all the same. The type determines your entire strategy.

Type 1: Standard Prediction (Tabular)

Examples: House Prices, TPS series, Spaceship Titanic, s6e7 Mechanism: Upload CSV file. Scored immediately on fixed test set. Scoring: Deterministic (RMSE, AUC, logloss — same formula every time) Public kernel value: ★★★★★ (best public kernel ≈ 95-99% of winning score) GPU needed: Usually no (CPU GBDT sufficient) Quota: 5/day, no convergence delay

Strategy:

1. Day 1: Fork best public kernel → submit → establish baseline
2. Day 2-3: AutoGluon best_quality → compare with public → blend if complementary
3. Day 4-7: Feature engineering (highest ROI lever for tabular)
4. Day 8+: Model diversity (LGB+XGB+CAT blend), pseudo-labeling (cautious)
5. Final: Submit best blend, stop tuning 24h before deadline

Validation: 5-fold StratifiedKFold + adversarial validation. Trust OOF if N>10K.

Type 2: Code Competition (Notebook Required)

Examples: ROGII Wellbore, Biohub Cell Tracking, NeuroGolf Mechanism: Submit Kaggle Notebook. Runs on hidden test set. No internet. Scoring: Deterministic on hidden test (may differ significantly from public) Public kernel value: ★★★★☆ (fork is main strategy, but artifacts/dependencies) GPU needed: Often yes (UNet, LLM inference, model training) Quota: 5/day or 10/day, kernel run time limit (9-12h)

Strategy:

1. Day 1: Read top 10 public kernels. Identify required datasets (artifacts).
2. Day 2: Fork best kernel. Add ALL required datasets as inputs. Push → verify COMPLETE.
3. Day 3: If kernel ERRORs → read log → fix dependencies (see code-competition-artifact-pipeline)
4. Day 4-7: Modify pipeline (trim crashing components, tune parameters)
5. Final: Ensure kernel COMPLETES (ERROR → no score, even if submission.csv exists)

Critical: Code competitions do NOT score ERROR-status kernels. A kernel that crashes mid-pipeline produces NO submission, even if submission.csv was written before the crash. Always trim or try/except around risky cells.

Type 3: Simulation Competition (TrueSkill)

Examples: PTCG AI Battle, Orbit Wars, Connect-X Mechanism: Submit agent (tar.gz or notebook). Agent plays games against other agents on the ladder. TrueSkill Bayesian rating N(μ,σ²). Scoring: Probabilistic (μ changes with each game, σ decreases over time) Public kernel value: ★★★☆☆ (fork gets you started, but meta shifts weekly) GPU needed: Usually no (game agents are rule-based or RL) Quota: 5/day, but DON'T use them all — see below

Strategy (CRITICAL — read trueskill-simulation-competition-strategy):

1. Day 1-3: Fork best public agent. Build local eval harness (free CPU kernel).
2. Day 4-7: Submit 1-2 agents. DO NOT re-submit to "test" (resets μ to 600).
3. Week 2-4: Let agents converge (σ decreases). Monitor LB but don't act on noise.
4. Week 5+: Near deadline, submit 2 duplicate copies of best agent (high-roll).
5. Final 2 weeks: Kaggle continues running games until convergence.

Key rule: Latest 2 submissions count for final. ALL agents keep playing. LB shows best-ever score, but FINAL uses latest-2 only.

Type 4: Research Competition

Examples: Biohub Cell Tracking, ARC Prize Mechanism: Notebook submission + sometimes written report Scoring: Custom metric (often domain-specific) Public kernel value: ★★☆☆☆ (less community, harder problems) GPU needed: Usually yes (heavy compute for research-grade tasks) Quota: 5/day, long timeline (3+ months)

Strategy:

1. Read the domain literature (papers, GitHub repos)
2. Fork best public kernel (often the only option — custom work is months of effort)
3. Focus on pipeline correctness (format, dependencies) before optimization
4. Allocate GPU wisely (30h/week shared across ALL competitions)

Type 5: Playground Series

Examples: s6e1-s6e7 (monthly) Mechanism: Standard prediction, synthetic data, swag prizes (not medals) Scoring: Deterministic (same as Type 1) Public kernel value: ★★★★★ (highly collaborative, best public ≈ winning) GPU needed: No (synthetic data is small/medium) Quota: 5/day, low stakes

Strategy:

1. Day 1: AutoGluon best_quality (5-15 min baseline)
2. Day 2: Fork best public → blend with AutoGluon
3. Day 3-5: Feature engineering, model diversity
4. Day 6+: Stop when ceiling reached (3-strike rule)
5. No stress — these are practice competitions

Type 6: LLM/Agent Benchmark

Examples: Industrial Automation Track 1, ARC-AGI-3, ai-agent-security Mechanism: Submit notebook with model/prompt. Evaluated on hidden prompts. Scoring: Accuracy or custom metric Public kernel value: ★★☆☆☆ (LLM space moves fast, public kernels expire) GPU needed: Yes (running open-source LLMs) Quota: 5-10/day, GPU-intensive

Strategy:

1. Choose the right open-source model (Qwen, DeepSeek, LLaMA — competition rules)
2. Prompt engineering is 80% of the work
3. Few-shot examples from training data
4. Closed-book setting: no RAG, no tools — rely on model's internal knowledge
5. Test locally with vLLM before submitting

The First-48-Hours Decision Tree

New Competition Found
│
├── Read Overview + Evaluation + Rules (30 min)
│
├── Identify Type
│   ├── Standard Tabular → AutoGluon first, then fork public
│   ├── Code Competition → Find artifacts, fork best public
│   ├── Simulation → Read TrueSkill strategy, build eval harness
│   ├── Research → Read literature, fork best public
│   ├── Playground → AutoGluon, blend, practice
│   └── LLM Benchmark → Choose model, prompt engineer
│
├── Scan Top 10 Public Kernels
│   ├── Sort by score descending
│   ├── Read their approach (title + first cell)
│   ├── Check: does best public use external data? special tricks?
│   └── Estimate: 0.6×BEST_PUBLIC baseline
│
├── Submit Best Public Kernel (1 quota)
│   ├── Validates format + baseline
│   └── If LB < 0.6×BEST_PUBLIC → something is wrong (check format)
│
└── Plan Quota Allocation
    ├── Week 1: 2-3 submissions (baseline + first improvement)
    ├── Week 2-3: 1-2 submissions/week (only verified improvements)
    └── Final week: 1-2 submissions (high-roll for simulation, final blend for tabular)

Quota Allocation Framework

For Standard/Playground (5/day)

Total submissions available: 5 × (days_until_deadline)
Recommended usage:
  - 30% baseline + public fork exploration
  - 40% verified improvements (OOF + format check before submit)
  - 20% ensemble/blend experiments
  - 10% reserved for emergencies/final day

For Simulation (5/day, but DON'T use)

Total submissions available: 5 × (days_until_deadline)
Recommended usage:
  - Week 1: 1-2 (establish agent, start convergence)
  - Week 2-N: 0 (let convergence happen, don't re-submit)
  - Final 2 weeks: 2 (duplicate best agent for high-roll)
  - NEVER: re-submit same agent to "test" (resets convergence)

For Code Competition (5/day)

Total submissions available: 5 × (days_until_deadline)
Recommended usage:
  - Day 1-3: 2-3 (fork + fix dependencies + first COMPLETE run)
  - Day 4-7: 2-3 (pipeline optimization, parameter tuning)
  - Week 2+: 1/week (only significant improvements)
  - Final: 1-2 (submission format verification + final model)

GPU Quota Management (30h/week shared)

GPU time is the scarcest resource. It's shared across ALL your competitions.

Priority allocation:
1. Active competition with deadline < 2 weeks → 50% GPU
2. Active competition with deadline < 1 month → 30% GPU
3. Exploration/baseline runs → 15% GPU
4. Reserved for emergencies → 5% GPU

Anti-pattern: running AutoGluon best_quality (8h) on a Playground competition while a Code competition deadline is 3 days away.

Evidence

Validated across 20+ competitions:

• Standard: House Prices (0.11750), SST (0.80780), s6e7 (0.94942), TPS May (0.99754)
• Code: ROGII (Pipeline A 10.38), Biohub (v1 fork), NeuroGolf (7228)
• Simulation: PTCG (LB 967.8, rank 219/4164), Orbit Wars (observed)
• Research: ARC-AGI-3 (Duck harness studied)
• Playground: s6e1-s6e7 (6 competitions, mean top 5%)
• LLM: Industrial Automation Track 1 (FMEA knowledge base)