These questions span the full stack of RL for trading: from theoretical foundations to practical implementation. They are intentionally broad — we believe the highest-impact work sits at the intersection of multiple disciplines.
Fundamental Questions
Q1. Can RL consistently beat buy-and-hold on Indian equities after realistic transaction costs?
Current evidence: mixed. SAC achieves Sharpe 0.8-1.2 on test data, but buy-and-hold NIFTY50 averaged ~12% annually (2015-2024) with minimal costs. The bar is higher than it looks.
Q2. What is the minimum signal-to-noise ratio required for RL to learn profitable policies?
NSE daily data has SNR < 0.05. Is this sufficient, or does RL require higher-frequency data (minute bars) to find exploitable patterns?
Q3. Can a single policy generalize across all market regimes, or does regime-switching require explicit meta-learning?
Current single-regime policies degrade 15-30% across regime transitions. Can we build a policy that is inherently robust, or do we need a meta-learner that selects regime-specific sub-policies?
Q4. What is the optimal action space for portfolio management?
Discrete K-slot (current) vs. continuous target weights vs. hierarchical buy/sell/size? Each has trade-offs in expressiveness, sample efficiency, and interpretability.
Q5. How much does survivorship bias affect RL training on Indian equities?
We only train on stocks that still exist. How much would including delisted stocks change performance? Can we model the selection process explicitly?
Algorithm Design Questions
Q6. Would offline RL outperform online RL for trading?
Offline RL (CQL, IQL, Decision Transformer) avoids the sim-to-real gap by learning from historical data directly. Does this help or hurt given the limited data scale?
Q7. Can curriculum learning accelerate convergence?
Start with simplified environments (no transaction costs, single stock) and gradually increase complexity. Does the agent learn faster, or does it overfit to the curriculum?
Q8. Would model-based RL be more sample-efficient?
Learn a dynamics model of price evolution and plan within imagined trajectories. Financial time series are notoriously hard to model — can MBPO or Dreamer help?
Q9. Can multi-agent RL capture adversarial market dynamics?
Model other market participants as agents competing for alpha. Does this improve robustness, or does it introduce instability?
Q10. Would distributional RL better capture tail risks?
The return distribution is heavy-tailed; mean Q-values miss the tail risk. Can QR-DQN or IQN improve risk-adjusted performance?
Feature & Data Questions
Q11. Would alternative data add alpha?
News sentiment, social media, satellite imagery — currently unused. Would any of these improve NSE-specific performance given data quality and latency constraints?
Q12. Can attention mechanisms over the stock universe improve cross-sectional selection?
Current features treat stocks independently. Can a transformer encoder model inter-stock relationships for better relative-value selection?
Q13. Would graph neural networks capture inter-stock relationships?
Sector membership, supply chain links, and correlation structure as graph edges. Does this improve diversification and risk management?
Q14. How much value does intraday data add over daily bars?
Current research focuses on daily data. Minute data has 375x more bars per day but also higher noise and infrastructure requirements. Where is the sweet spot?
Q15. Can time-series foundation models provide useful embeddings?
Chronos, TimesFM, Lag-Llama — pre-trained on diverse financial data and fine-tuned for NSE. Initial experiments show marginal improvement. Is the integration suboptimal?
Risk & Portfolio Questions
Q16. How should position sizing interact with the RL policy?
Currently, the agent chooses which stock to trade but not how much. Should sizing be part of the action space, or should a separate risk layer handle it?
Q17. Can risk parity be learned end-to-end?
Rather than applying risk management as a post-hoc constraint, can the agent internalize equal-risk-contribution allocation?
Q18. Would options strategies improve risk-adjusted returns?
NSE F&O market is liquid enough for protective puts and covered calls. Can the agent learn to use derivatives for hedging?
Q19. Can the agent learn to dynamically adjust its risk appetite based on regime detection?
Currently, risk profiles are fixed per episode. Adaptive profiles within an episode could better respond to changing market conditions.
Q20. What is the optimal stop-loss strategy for RL-managed positions?
Current hard stop-loss may conflict with the agent's learned policy. Should stops be part of the reward structure, or should the agent learn its own exit logic?