AlgoGPT: Open-Source AI Model for Indian Derivatives and Equity Market Data

AlgoGPT is the first open-source model for financial candlesticks (K-lines). Built with Kronos open-source foundation model. Trained on data from NSE & BSE.

Core Algorithem:

• [Release] Jun, 2024: Released AlgoGPT-Broker Setup Module! 🔥Base Model, Core Structure!
• [Release v2.0] October, 2026: Upgraded with Kronos Foundation Model (BSQ K-Line Tokenizer + Causal Autoregressive Transformer), multi-symbol walk-forward backtesting, and automated market data feeds (Yahoo Finance, Upstox, NSE).

Model Performace:

• [Version 1.0] April, 2024: Base Model Model Accuracy: 0.43
• [Version 2.0] October, 2026: Kronos Foundation Engine (NeoQuasar/Kronos-small & Kronos-Tokenizer-base) with continuous BSQ candlestick tokenization, multi-step autoregressive price path generation, and systematic options derivative backtesting with risk metrics (Sharpe Ratio, Alpha, Max Drawdown).

Architecture

algogpt/
├── model/           # Kronos foundation model (tokenizer, model, predictor)
├── data/            # Data fetching & dataset creation for Indian derivatives
├── training/        # Finetuning pipelines (tokenizer + predictor)
├── trading/         # Feature engineering, signals, backtesting
├── main.py          # Main trading pipeline entry point
├── quickstart.py    # Quick test with Yahoo Finance
└── requirements.txt # Dependencies

Quick Start

# Install dependencies
pip install -r requirements.txt

# Run quick test (uses Yahoo Finance for NIFTY daily data)
python quickstart.py

Data Sources

Finetuning on Indian Derivatives

1. Prepare Data

# Download NIFTY/BANKNIFTY 5min data to ./data/
# Or use Upstox API with credentials

2. Configure Training

Edit training/config.py or use environment variables:

export SYMBOLS="NIFTY,BANKNIFTY"
export DATA_SOURCE="csv"
export DATA_DIR="./data"
export INTERVAL="5min"
export START_DATE="2020-01-01"
export END_DATE="2024-12-31"
export PREDICTOR_EPOCHS=30
export BATCH_SIZE=50

3. Run Distributed Finetuning

# Tokenizer finetuning (run first)
torchrun --standalone --nproc_per_node=1 training/train_tokenizer.py

# Predictor finetuning (uses finetuned tokenizer)
torchrun --standalone --nproc_per_node=1 training/train_predictor.py

For multi-GPU:

torchrun --standalone --nproc_per_node=4 training/train_predictor.py

4. Use Finetuned Models

from algogpt.model import KronosPredictor

predictor = KronosPredictor.from_local(
    model_path="./outputs/models/finetune_predictor_indian/checkpoints/best_model",
    tokenizer_path="./outputs/models/finetune_tokenizer_indian/checkpoints/best_model",
)

Trading Pipeline

# Basic backtest on NIFTY
python main.py --symbol NIFTY --start 2023-01-01 --end 2024-06-01

# With Upstox data
python main.py --symbol BANKNIFTY --data-source upstox \
  --upstox-api-key YOUR_KEY --upstox-api-secret YOUR_SECRET \
  --upstox-access-token YOUR_TOKEN

# Using finetuned model
python main.py --model ./outputs/models/finetune_predictor_indian/checkpoints/best_model \
  --tokenizer ./outputs/models/finetune_tokenizer_indian/checkpoints/best_model

Key Components

Model (algogpt/model/)

• KronosTokenizer: Hybrid quantization (BSQ) encoder-decoder
• Kronos: Autoregressive transformer with hierarchical tokens (s1+s2)
• KronosPredictor: High-level inference with normalization, sampling, batch prediction

Data (algogpt/data/)

• DataFetcher: Abstract base + Upstox/NSE/Yahoo/CSV implementations
• IndianDerivativesDataset: Sliding window dataset with proper normalization
• create_train_val_datasets: Time-aware train/val split (no lookahead bias)

Training (algogpt/training/)

• FinetuneConfig: Centralized dataclass config with env overrides
• train_tokenizer.py: DDP tokenizer finetuning (BSQ loss)
• train_predictor.py: DDP predictor finetuning (next-token CE loss)
• utils.py: DDP setup, seeding, checkpointing, Comet logging

Trading (algogpt/trading/)

• FeatureEngineer: 30+ technical indicators (RSI, MACD, BB, candles, etc.)
• SignalGenerator: Kronos + TA weighted ensemble with MACD filter
• KronosBacktester: Walk-forward with realistic execution (next-bar open, slippage, commission)

Indian Derivatives Specifics

Time features: minute, hour, weekday, day, month (IST timezone)

Configuration

All configs use Python dataclasses with:

• Type hints and defaults
• Environment variable overrides (export PARAM=value)
• Validation in __post_init__
• Dict-like access for backward compatibility

Requirements

• Python 3.10+
• PyTorch 2.0+ (CUDA recommended)
• 8GB+ VRAM for finetuning (batch_size=50)
• 16GB+ RAM for data processing

License

MIT License - See LICENSE file for details.