Kaggle Store Sales --- Capability Formation Case

Language / 语言: English primary · 中文概览如下。

中文概览

使用公开时间序列竞赛作为可度量训练环境的纵向学习项目，沉淀预测实验和可复用机器学习知识。

A longitudinal learning project for TopPrism's Cultivating ML Agent, using a public time-series competition as a measurable training environment.

LEARNING PROJECT · TIME SERIES · KAGGLE BENCHMARK

This is not a customer product. Its value is the project experience, failure analysis, and reusable ML knowledge crystallized for later work.

---

Why this project matters

The public competition provides:

• a fixed dataset;
• a measurable metric;
• temporal validation challenges;
• feature engineering;
• model comparison;
• operational debugging.

That makes it useful as an environment where an ML agent can learn and leave behind reusable skills.

---

Competition task

Forecast 16 days of sales for 54 stores × 33 product families using the public Favorita Store Sales dataset.

---

Most important learning

The repository should emphasize what became reusable knowledge, not every implementation detail.

Examples already visible in the project:

• lag / rolling features must be shifted to avoid leakage;
• temporal validation should use rolling / walk-forward logic;
• memory behavior in large pandas feature pipelines matters;
• public-LB failure can reveal train/test feature-generation mismatch;
• external knowledge sources are useful only when validated against the actual pipeline.

---

Preserved failure history

The project records a large CV vs public-LB mismatch and the iterative fixes that followed. That history is preserved verbatim in docs/competition-log.md because Native AI needs to learn from failure, not only from final scores.

The reusable pattern in this project is:

Project
-> failure
-> root cause
-> corrected method
-> reusable skill / principle
-> later reuse

The v1 -> v3 (LightGBM ffill fix) and v3 -> v4 (Round 2 multi-model ensemble) jumps in this project both follow this pattern.

---

Evidence

Competition scores are classified as learning evidence, not as product evidence. Long feature-by-feature tutorials and code-debug narratives live in docs/competition-log.md and are not duplicated in the README.

This README records:

• task: forecast 16 days of sales for 54 stores x 33 product families;
• best reproducible result: LightGBM with ffill fix (public LB ~1.90248 per the preserved log);
• most important failures: v1 -> v2 NaN-vs-zero lag-fill regression;
• skills crystallized: time-series walk-forward validation, feature-leakage detection, AutoML-first tabular baseline;
• relationship to cultivating-ml-agent: this is one of the longitudinal capability-formation cases the agent ran through.

---

TopPrism metadata

topprism:
  purpose: learning-project
  capability: time-series-ml
  maturity: learning
  evidence:
    type: kaggle-benchmark
  parent:
    - cultivating-ml-agent