NSE PORTFOLIO
● Paper Trading Phase NSE · BSE

Reinforcement Learning
for Indian Markets

An institutional-grade AI trading platform for NSE & BSE. Rigorously validated — not just backtested. Currently in research. Not claiming live alpha yet.

NSE Stocks Covered
Time Granularities
Data Files
Market Indices
Representative Test Period Equity Curve +15.3%
Illustrative single episode from multi-episode test evaluation
The Problem

Algo Trading is Hard.
Most Fail in Production.

Building a strategy that looks great in backtests is easy. Making it survive live markets is not.

01

Curve Fitting

A model optimized on historical data captures noise, not signal. Without walk-forward validation, backtest performance is fiction.

02

Unrealistic Simulation

Ignoring NSE transaction taxes (STT, stamp duty, exchange fees), slippage, and minimum lot sizes inflates paper returns by 30–50%.

03

Slow Iteration

Training RL agents for days per experiment destroys research velocity. You can't build intuition about what works without rapid feedback.

04

No Risk Model

Agents that maximize raw returns without drawdown constraints become reckless. Risk must be part of the objective, not an afterthought.

What We Built

Six Core Capabilities
That Address Each Failure Mode

RL-Powered Agents

Deep reinforcement learning agents trained on real NSE data with realistic transaction costs and risk constraints.

Actor-Critic · JAX · PyTorch Multi-stock portfolio

Walk-Forward Validation

Multi-window rolling validation. The agent must generalize across time periods — not just memorize one regime.

Rolling windows Cross-regime consistency

GPU-Accelerated Training

XLA-compiled training loops for rapid iteration. Experiments complete in minutes, not days.

>9K steps/sec Consumer GPU

NSE-Realistic Simulation

Full Indian brokerage cost model: STT, exchange fees, SEBI charges, GST, stamp duty. Slippage and market impact included.

Full NSE fee structure Broker API integration

Multi-Granularity Data

Extensive coverage across timeframes and indices. From minute bars to daily, spanning large-cap to small-cap universes.

Multi-timeframe NIFTY 50 to Smallcap

Risk-Aware Reward Design

Multi-objective reward function combining returns, drawdown penalties, concentration limits, and turnover costs. Agent learns risk discipline, not just return maximization.

Multi-component Risk-adjusted
Live Data

Trained on Real NIFTY Constituents

Not synthetic data. Historical data fetched directly via broker API, stored in columnar format with quality validation.

700+Stocks Tracked
5000+Data Files
8Granularities
6Indices Covered
How It Works

Six-Stage Pipeline
Every Stage Gated

An agent only advances when it clears measurable criteria. No exceptions. This is what separates research-grade from production-grade.

Stage 01Complete

Data Ingestion

Extensive NSE/BSE coverage across multiple granularities via broker API. Checkpointed parallel downloads, quality validation, multi-source fallback.

Stage 02Complete

Feature Engineering

Technical indicators and auxiliary signal models for regime detection, volatility forecasting, and trend classification.

Stage 03Complete

RL Agent Training

Agents trained with NSE-realistic frictions and strict promotion gates. Only agents that clear all criteria advance to the next stage.

Stage 04In Progress

Walk-Forward Validation

Multi-window rolling test proving the agent generalizes — not just memorizes one period. Adversarial stress tests under extreme market conditions.

Stage 05In Progress

Paper Trading

30+ day live simulation against real market data. Realistic order fills, portfolio tracking, real-time P&L. An agent that cannot pass paper trading does not go live.

Stage 06Planned

Live Deployment

Low-latency inference via broker API. Live risk controls: position limits, drawdown circuit breakers, concentration guards. Real capital only after validation and paper trading clear.

Performance

Real Experiment Results.
Not Illustrations.

These numbers come directly from experiment records. Test metrics are on held-out data the agent never saw during training. Past simulation results do not guarantee future live performance.

Test Sharpe
Risk-adjusted return on held-out test data.
Test Win Rate
% of test episodes with positive return.
Val Sharpe
Validation set used for checkpoint selection.
Training Speed
Steps per second on GPU with JAX XLA.

Illustrative Equity Curve — Representative Test Episode

Generated using real test statistics from held-out evaluation.

● RL Portfolio ⋯ NIFTY 50 (est.)

Curve parameterised from real test statistics, not cherry-picked. Benchmark uses estimated NIFTY 50 historical parameters. This is one representative episode from multi-episode evaluation — not the best one.

Honest Status

Where We Actually Are

Transparency is a feature. Here is the current development state, without spin.

What's Done

  • Full data pipeline for Indian equities
  • Dual RL training backends for research flexibility
  • Validated training pipeline with reproducible results
  • NSE-realistic transaction cost model
  • GPU-accelerated training infrastructure
  • Research dashboard for monitoring and analysis
  • CLI tooling for data, training, and backtesting
  • Initial backtesting suite across multiple strategies

In Progress

  • Walk-forward validation across multiple windows
  • Adversarial stress testing framework
  • Paper trading (live simulation, no real capital)
  • Unified architecture across training backends
  • Curriculum learning for exploration
  • Expanded universe coverage

Planned

  • Live deployment via broker API (post paper trading gate)
  • Automated hyperparameter search with risk adaptation
  • Multi-agent portfolio coordination
  • Institutional API / white-label deployment
  • Regulatory compliance pathway

We are a research-stage project. No real capital is deployed. We are seeking collaborators and investors who understand the difference between a validated research prototype and a live product — and who see the value in what's been built so far.

Research Log

Experiment History
Every Run Logged

Real experiment IDs, real timestamps, real metrics from disk. Every run is reproducible from its checkpoint.

Selected metrics from representative training runs. All results on held-out test data the agent never saw during training.

Next Experiments

Active Walk-forward validation · Adversarial stress testing · Curriculum learning
Planned Multi-seed statistical validation · Minute-bar pilots · NIFTY 100 expansion
Read Full Experiment Write-ups →
Technology

Built on Production-Grade
Open-Source Foundations

Py
PythonCore language
JAX
Google JAXXLA compilation
PT
PyTorchGPU training
Gym
GymnasiumRL environment interface
ZK
Broker APILive market access
AN
AnalyticsData exploration
CU
CUDAGPU acceleration
PQ
Parquet + SQLiteHybrid storage
About

The Builder

Solo Researcher & Engineer

RL / Quantitative Systems · Indian Markets

Building production-grade reinforcement learning systems for Indian equities from first principles. Background spans machine learning research, systems engineering, and quantitative analysis. This project exists because RL for Indian markets is severely underexplored — and because most "AI trading" products are backtested fiction. I believe rigorous validation, honest metrics, and realistic cost models are the only path to real alpha.

Substantial
Modern research codebase
1
Published experiment result (more in pipeline)
6
Phase development roadmap
5000+
Real market data files on disk
Who This Is For

Four Paths to Engagement

01

Investors

Early-stage opportunity in a validated RL trading research platform built for Indian equities. Pipeline is clear. Risk gates are real. Results published, more in pipeline.

→ Request research materials and methodology
02

Quant Researchers

Looking for a serious collaborator on RL for Indian markets? The infrastructure is built. The hard problems remaining are algorithmic: walk-forward consistency, multi-agent coordination, curriculum learning.

→ Discuss research collaboration on specific open problems
03

ML / RL Engineers

Interested in contributing to a production-bound RL system? Phases 3–6 of the roadmap involve architecture consolidation, AutoRL, and NIFTY 100+ universe expansion.

→ Review the open roadmap and reach out with your background
04

FinTech / Institutions

Need an AI trading engine for Indian equities that you can run in-house? This is modular: swap broker backends, risk profiles, and asset universes. Licensing and white-label options available post-Stage 5.

→ Discuss licensing and integration after paper trading gate
Contact

Let's Talk.
Seriously.

No pitch decks sent cold. If you've read this far, you understand what's been built and what stage it's at. That's the kind of conversation worth having.

Investors Research materials and detailed methodology available under NDA.
Collaborators Tell me what problem you want to work on. Open problems listed in the research section.
FinTech / Institutional Licensing discussions welcome post-Stage 5. Let's establish a relationship now.

Strictly confidential. No spam, ever.