PokieTicker — Understand the "Why" Behind Every Price Move

🔗 Live Demo: mitrui.com/PokieTicker |

Since I wanted to understand the stories behind candlestick charts, I vibe coded this app.

As a stock market beginner, the news was too fragmented and I was wasting a lot of time, never understanding why prices went up or down. I built PokieTicker to develop event-driven thinking — to understand the "why" behind every price move.

Click to watch the full demo video

What It Does

• News on the chart — Dots on the candlestick chart represent news for each date. Click any dot to see the news impacting that company at the time.
• Filter by impact type — Market, earnings, product, policy, competition, or management. Click a category to see related bullish or bearish news.
• Find similar events — Discover historical days with similar news patterns and see what happened to the stock afterward.
• AI explains price moves — Select a date range and ask AI why the stock dropped or rallied. It tells you what events caused it.
• Predict trends — Based on the past 30 days of news events, the model predicts how these events might affect future price direction.

How the Prediction Works

PokieTicker includes an XGBoost-based prediction system that combines news sentiment with technical indicators:

Features (31 total):

• News features — article count, sentiment score, positive/negative ratio, 3/5/10-day rolling averages, sentiment momentum
• Technical features — price returns (1/3/5/10-day), volatility, volume ratio, RSI-14, moving average crossover

How it works:

1. News articles are scored for sentiment by Claude Haiku (batch API) — each article gets a sentiment label, key discussion summary, and bullish/bearish reasons
2. These are combined with OHLC price data into daily feature vectors
3. XGBoost classifiers predict up/down direction at T+1, T+3, and T+5 horizons
4. The system also finds historically similar periods (cosine similarity on feature vectors) and shows what happened next

Why news-driven prediction can work:

• Stock prices are driven by information. Major news events (earnings, policy changes, product launches) create predictable short-term momentum
• Sentiment clustering matters — when multiple negative articles cluster on the same day, the downward pressure tends to persist for 1-3 days
• Historical pattern matching works because markets react similarly to similar events (e.g., tariff announcements, FDA approvals, earnings beats)

This is an experimental tool for learning, not financial advice. Markets are complex and no model captures everything.

Architecture

Frontend (React + Vite + D3.js)          Backend (FastAPI + SQLite)
+---------------------------------+      +----------------------------+
|  CandlestickChart (D3.js)       |      |  /api/stocks/{sym}/ohlc    |
|  +- news dots on each date      |----->|  /api/news/{sym}?date=     |
|  +- crosshair + click to lock   |      |  /api/news/{sym}/categories|
|                                  |      |                            |
|  NewsPanel (right sidebar)       |<-----|  SQLite: pokieticker.db    |
|  +- sentiment sorted             |      |  +- ohlc (51K+ rows)      |
|  +- up/down reasons              |      |  +- news_raw (61K+)       |
|  +- T+1/T+5 returns              |      |  +- layer1_results (97K+) |
|                                  |      |                            |
|  PredictionPanel                 |<-----|  /api/predict/{sym}/forecast|
|  +- 7-day & 30-day forecasts     |      |  +- XGBoost models         |
|  +- similar historical periods   |      |  +- cosine similarity      |
+---------------------------------+      +----------------------------+

Data Pipeline

Polygon API --> Layer 0 (rule filter) --> Layer 1 (Haiku Batch API) --> Layer 2 (Sonnet, on-demand)
  OHLC + News    reject spam/listicles    sentiment + up/down reasons     deep analysis on click
                  ~17% rejected            50 articles per batch call      cached in DB

Quick Start

The repo includes a pre-built database (pokieticker.db) with historical data, so you can run it immediately — no API keys needed.

git clone https://github.com/owengetinfo-design/PokieTicker.git
cd PokieTicker

# Unpack the pre-built database
gunzip -k pokieticker.db.gz

# Backend (Python 3.10+)
python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -r requirements.txt

# Frontend (Node.js 18+)
cd frontend && npm install && cd ..

Then start both services (in two terminal windows):

# Terminal 1: Backend
source venv/bin/activate
uvicorn backend.api.main:app --reload

# Terminal 2: Frontend
cd frontend && npm run dev

Open http://localhost:7777/PokieTicker/ and you're done.

Just want to see the demo? Visit mitrui.com/PokieTicker — no setup needed.

Updating data (optional)

If you want to fetch the latest stock data and run AI analysis, you'll need API keys:

cp .env.example .env
# Edit .env and fill in your keys:

Key	Where to get it	Cost
POLYGON_API_KEY	polygon.io	Free tier
ANTHROPIC_API_KEY	console.anthropic.com	Pay-as-you-go

# Fetch new OHLC + news
python -m backend.bulk_fetch

# Run AI analysis
python -m backend.batch_submit --top 50
python -m backend.batch_collect <batch_id>

Project Structure

.env                          # API keys (gitignored)
pokieticker.db                # SQLite database (gitignored)
requirements.txt              # Python dependencies

backend/
  config.py                   # pydantic-settings, loads .env
  database.py                 # 9-table SQLite schema
  bulk_fetch.py               # Bulk download OHLC + news for many tickers
  batch_submit.py             # Submit Layer 1 to Anthropic Batch API
  batch_collect.py            # Collect Batch API results
  weekly_update.py            # Incremental weekly data update
  polygon/
    client.py                 # Polygon API with retry/backoff
  pipeline/
    layer0.py                 # Rule-based filter (free)
    layer1.py                 # Claude Haiku batch analysis
    layer2.py                 # Sonnet on-demand deep analysis
    alignment.py              # News -> trading day + forward returns
    similarity.py             # Similar news pattern matching
  ml/
    features.py               # 31-feature engineering (news + technical)
    model.py                  # XGBoost training (per-ticker + unified)
    inference.py              # Forecast generation + similar period analysis
    backtest.py               # Backtesting framework
    lstm_model.py             # Experimental LSTM model
  api/
    main.py                   # FastAPI app + CORS
    routers/
      stocks.py               # GET /api/stocks, /search, /{sym}/ohlc
      news.py                 # GET /api/news/{sym}, /{sym}/timeline
      analysis.py             # POST /api/analysis/deep, /story
      pipeline.py             # POST /api/pipeline/fetch, /process
      predict.py              # GET /api/predict/{sym}/forecast

frontend/
  src/
    App.tsx                   # Main layout (chart + panels)
    components/
      CandlestickChart.tsx    # D3.js chart with news dots
      NewsPanel.tsx           # Sentiment-sorted news cards
      NewsCategoryPanel.tsx   # Filter by impact category
      PredictionPanel.tsx     # AI forecast + similar periods
      RangeQueryPopup.tsx     # "Why did it drop?" range analysis
      SimilarDaysPanel.tsx    # Historical pattern matches
      StockSelector.tsx       # Ticker search + tabs

Weekly Update

# Fetch new OHLC + news since last update
python -m backend.weekly_update

# Run AI analysis on new articles
python -m backend.batch_submit --top 50
python -m backend.batch_collect <batch_id>

Cost Summary

Item	Cost
Polygon data (free tier)	$0
Layer 1 Batch API (per 1000 articles)	~$0.35
Layer 2 on-demand (per article)	~$0.003
Weekly incremental update	~$1-2

Tech Stack

• Frontend: React, TypeScript, Vite, D3.js
• Backend: FastAPI, SQLite (WAL mode), Pydantic
• AI: Claude Haiku 4.5 (batch sentiment), Claude Sonnet (deep analysis)
• ML: XGBoost (prediction), cosine similarity (pattern matching)
• Data: Polygon.io REST API

License

MIT — see LICENSE for details.