# Ml Classification Project

> Build the ML-classification portfolio project: engineer technical-indicator features, train an XGBoost model to predict whether the S&P 500 will be up or down tomorrow, benchmark it against a coin flip, and publish a GitHub repo.

- Skill: `knuckles-team/ml-classification-project` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add knuckles-team/ml-classification-project`
- Raw SKILL.md: https://api.skillmd.com/api/skills/knuckles-team/ml-classification-project/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: Knuckles-Team (https://skillmd.com/u/knuckles-team)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/knuckles-team/ml-classification-project

---


# ML Classification Project Workflow

Project 5 of the quant portfolio. Predict next-day direction (not price) and
benchmark honestly against a coin flip. See
`quant-career-docs/reference/ml-for-finance.md`.

## Steps

### Step 1: engineer-features
**Agent**: `quant_researcher`
**Tools**: `data-science-mcp`

Pull S&P 500 history via yfinance and engineer technical-indicator features
(returns, moving averages, RSI, volatility, momentum). Ensure features use only
past data — no lookahead.
Expected: `feature-matrix`

### Step 2: train-xgboost [depends_on: Step 1]
**Agent**: `quant_researcher`

Train an XGBoost/LightGBM classifier to predict up/down for the next day. Use
time-aware train/test splitting (no shuffling across time).
Expected: `trained-classifier`

### Step 3: benchmark-coinflip [depends_on: Step 2]
**Agent**: `quant_researcher`

Evaluate out-of-sample accuracy/AUC and benchmark against a 50/50 coin flip and a
naive majority-class baseline.
Expected: `benchmark-comparison`

### Step 4: evaluate [depends_on: Step 3]
**Agent**: `risk_analyst`

If the classifier drives a strategy, report Sharpe, max drawdown, CAGR vs
benchmark. State whether the edge survives costs — honest analysis.
Expected: `evaluation-metrics`

### Step 5: github-publish [depends_on: Step 4]
**Agent**: `quant_developer`

Publish a GitHub repo with README, metrics, the coin-flip benchmark, clean code,
and an honest "what didn't work" section.
Expected: `github-repo-url`

### Step 6: kg-persist [depends_on: Step 5]
**Agent**: `quant_researcher`
**Tools**: `graph_write`

Persist the project and metrics as typed nodes linked to the portfolio.

## Output
- An XGBoost direction classifier benchmarked against a coin flip
- A published GitHub portfolio repo

## Execution

Run this workflow as a dependency-ordered DAG. Steps with no unmet `depends_on` run in parallel; dependents run after their prerequisites complete.

- **Run first (in parallel):** Step 1 — engineer-features
- **After level 0:** Step 2 — train-xgboost
- **After level 1:** Step 3 — benchmark-coinflip
- **After level 2:** Step 4 — evaluate
- **After level 3:** Step 5 — github-publish
- **After level 4:** Step 6 — kg-persist

**Execution:** If graph-os is reachable, offload the whole DAG via `graph_orchestrate action=execute_workflow` (or the `kg-delegate` skill) for true parallel/swarm execution. Otherwise execute the steps natively in dependency order: run steps with no unmet `depends_on` in parallel, then their dependents.