BEI Stock Predictor

Stock price prediction pipeline for Bursa Efek Indonesia (BEI/IDX) using three complementary models:

Model	Target	Script
LSTM	Next-day Close price (IDR)	lstm_predictor.py
Ridge Regression	Next-day return magnitude (%)	ridge_predictor.py
Logistic Regression	Next-day direction (up/down)	logistic_classifier.py

Data is sourced from Yahoo Finance and stored in a local SQLite database.

Features

• Download daily OHLCV data for any BEI/IDX ticker via Yahoo Finance
• Three independent prediction models that complement each other
• Combined predictor — one table with consensus signal from all three models
• Per-ticker hyperparameter search and saved optimal configs
• Batch prediction across all watchlist tickers (sorted by MAPE/MAE)
• Backtest mode for all predictors
• SQLite storage with idempotent upserts

Prerequisites

• Python 3.12+
• pip and venv

Installation

git clone <repo-url>
cd prediksaham

python -m venv .venv
source .venv/bin/activate

pip install -r requirements.txt

Key dependencies: yfinance, pandas, numpy, tensorflow==2.16.2, scikit-learn, matplotlib.

Usage

Always activate the virtual environment first:

source .venv/bin/activate

1. Download Data

# Default: all watchlist tickers, last 5 years
python bei_stock_downloader.py

# Single ticker, last 30 days
python bei_stock_downloader.py --ticker BBCA --days 30

# Multiple tickers
python bei_stock_downloader.py --tickers BBCA TLKM GOTO --days 14

# All tickers in watchlist.txt
python bei_stock_downloader.py --file watchlist.txt --days 30

# By years
python bei_stock_downloader.py --ticker BBCA --years 5

Edit watchlist.txt to manage your default ticker list (one IDX code per line, # for comments).

2. LSTM — Price Prediction

# Uses saved config from lstm_configs.json automatically
python lstm_predictor.py --ticker BBCA

# Override params
python lstm_predictor.py --ticker BBCA --lookback 20 --epochs 150 --forecast 3

# Save this run's config
python lstm_predictor.py --ticker TLKM --lookback 22 --save-config

# Backtest: train once, roll-predict last N days
python lstm_predictor.py --ticker BBCA --backtest 30

# Batch predict all configured tickers
python lstm_batch_predictor.py
python lstm_batch_predictor.py --ticker BBCA          # single ticker
python lstm_batch_predictor.py --backtest 30          # backtest all
python lstm_batch_predictor.py --ticker BBCA --backtest 30

Output: ranked table printed to terminal + predictions_YYYY-MM-DD.csv + prediction plot saved to prediction_images/.

Hyperparameter Search (LSTM)

# Search optimal lookback for one ticker
python lstm_config_search.py --ticker BBCA
python lstm_config_search.py --ticker BBCA --start 3 --end 60

# Batch search across multiple tickers
python lstm_batch_config_search.py
python lstm_batch_config_search.py --tickers ANTM CLEO
python lstm_batch_config_search.py --force   # re-run even if config exists

Results saved to ticker_configs_research/{TICKER}_lookback_search.csv and .png. Best config is auto-saved to lstm_configs.json.

3. Ridge Regression — Return Magnitude

# Find optimal config (all watchlist, or specific ticker)
python ridge_config_search.py
python ridge_config_search.py --ticker DMAS

# Predict next-day return %
python ridge_predictor.py --ticker DMAS
python ridge_predictor.py --all               # all watchlist tickers
python ridge_predictor.py --all --detail      # show model coefficients per ticker

# Backtest (last 30 trading days)
python ridge_predictor.py --ticker DMAS --backtest 30

4. Logistic Regression — Direction Classifier

# Find optimal config for all tickers in watchlist
python logistic_config_search.py

# Predict next-day direction (up/down + confidence)
python logistic_classifier.py --ticker DMAS
python logistic_classifier.py --ticker DMAS --detail  # show model coefficients
python logistic_classifier.py --all                   # all watchlist
python logistic_classifier.py --all --backtest 30     # backtest

5. Combined Prediction (All Models)

Run all three models for every ticker in the watchlist and display results in one consolidated table, ranked by consensus signal strength and confidence:

python combined_predict.py                          # all watchlist tickers
python combined_predict.py --tickers BBCA ANTM DMAS # specific tickers
python combined_predict.py --no-lstm                # skip LSTM model
python combined_predict.py --no-ridge               # skip Ridge model
python combined_predict.py --no-logistic            # skip Logistic model

Output columns: Ticker, Last Close, LSTM Forecast, LSTM Chg%, Ridge Ret%, Logistic, Conf, Sinyal, Rekomendasi

• Sinyal — how many models agree (e.g. 3/3 = all three bullish/bearish)
• Rekomendasi — consensus label: BELI KUAT / BELI / NETRAL / JUAL / JUAL KUAT
• Table is split into three sections: ▼ JUAL (top) → ◆ NETRAL → ▲ BELI (bottom); within each section sorted by signal strength then confidence

6. View Raw Data

python stock_viewer.py BBCA
python stock_viewer.py BBCA --data 20

Running Tests

source .venv/bin/activate
pytest

Project Structure

prediksaham/
├── bei_stock_downloader.py       # Download OHLCV → SQLite
├── stock_viewer.py               # View last N rows for a ticker
├── utils.py                      # Shared helpers (watchlist, config I/O)
├── watchlist.txt                 # Default ticker list
├── bei_stocks.db                 # SQLite database (local, not committed)
│
├── lstm_predictor.py             # LSTM: train + predict next-day price
├── lstm_config_search.py       # LSTM: lookback hyperparameter search
├── lstm_batch_config_search.py   # LSTM: batch hyperparameter search
├── lstm_batch_predictor.py         # LSTM: batch prediction for all tickers
├── lstm_configs.json             # Saved optimal LSTM configs per ticker
│
├── ridge_predictor.py            # Ridge: predict next-day return %
├── ridge_config_search.py        # Ridge: hyperparameter search
├── ridge_configs.json            # Saved optimal Ridge configs
│
├── logistic_classifier.py        # Logistic: predict next-day direction
├── logistic_config_search.py     # Logistic: hyperparameter search
├── logistic_configs.json         # Saved optimal Logistic configs
│
├── combined_predict.py           # Run all three models, one combined table
│
├── ticker_configs_research/      # Lookback search outputs (CSV + PNG)
├── prediction_images/            # LSTM prediction plots
└── tests/                        # pytest test suite

Database Schema

Table daily_prices in bei_stocks.db:

Column	Type	Description
Ticker	TEXT	IDX ticker code (e.g. BBCA)
Date	TEXT	Trading date (YYYY-MM-DD)
Open/High/Low/Close	REAL	Price in IDR, split/dividend adjusted
Volume	INTEGER	Shares traded
GainLoss_IDR	REAL	Close − Open (intraday move)
GainLoss_Pct	REAL	(Close − Open) / Open × 100
DayReturn_Pct	REAL	Day-over-day return vs previous Close
IntraDay_Range	REAL	High − Low (intraday volatility proxy)
UpdatedAt	TEXT	Timestamp of last upsert

Uniqueness constraint: UNIQUE(Ticker, Date). All upserts use INSERT OR REPLACE.

LSTM Configuration Reference

Default hyperparameters (overridden per ticker by lstm_configs.json):

Parameter	Default	Description
LOOKBACK	48	Days of history per sequence
FORECAST	1	Days ahead to predict
LSTM_UNITS	64	Neurons per LSTM layer
NUM_LAYERS	2	Stacked LSTM layers
DROPOUT	0.2	Dropout rate
EPOCHS	300	Max training epochs
PATIENCE	25	Early stopping patience

Notes

• BEI tickers use bare 4-letter IDX codes (BBCA, not BBCA.JK). The .JK suffix is added internally for Yahoo Finance.
• Prices are in IDR (Indonesian Rupiah).
• DayReturn_Pct for the first row of any date range is always NaN — skip it in return calculations.
• BEI is closed on Indonesian national holidays; date gaps are normal.
• bei_stocks.db is a local artifact and is not committed to git.