Projects
Research infrastructure
The systems the research runs on. Each project carries its real status — work that has not been built is labelled as a direction, not described as though it exists.
Current work
- Active ResearchPython· since 2025
Intraday Options Market Microstructure Research Pipeline
Python infrastructure that turns raw index and options-chain data into event-aligned, statistically testable research output.
- NIFTY spot ingestion and cleaning at one-minute resolution
- Market-structure engine detecting break-of-structure and change-of-character events
- Options-chain open-interest reconstruction around each event timestamp
- Event-time alignment of put- and call-side positioning
- Association testing with multiple-comparison correction
- NIFTY 50
- Open Interest
- Reproducibility
- Python
The published study's reproduction code is public. It requires market-data archives that are not redistributed; the data contract is documented in the repository.
Repository· github.com(opens in a new tab) - Under DevelopmentPython · Pine Script· since 2025
Market Structure Engine
A mechanical break-of-structure and change-of-character detector, specified precisely enough to be re-implemented from the description alone.
- Deterministic swing and structure-break identification
- Configurable significance threshold rather than discretionary judgement
- Parallel Pine Script and Python implementations used as a cross-check
- Event output consumed directly by the research pipeline
- Structural Breaks
- Market Structure
- CompletedPython· since 2025
Transaction-Cost-Aware Condor Hedge
A multi-leg backtesting framework that marks every leg to its actual traded price path, used to ask how much of a hedge's apparent edge survives real execution costs.
- All four legs of the structure marked to their own actual traded 1-minute price path, rather than an assumed slippage
- 322 sessions, 73 weekly expiries (Jan 2025 – Apr 2026)
- In sample, the Donchian buy-stop hedge improved risk-adjusted results and sharply reduced the worst drawdown
- Out of sample: the drawdown reduction replicated; the return improvement did not
- Per-leg effective-spread estimation (Corwin–Schultz; Abdi–Ranaldo) in place of assumed slippage
- Out-of-Sample Validation
- Options Hedging
- Transaction Costs
Code not published. See also: Donchian Breakout Option Overlay, an independently designed condor + Donchian-hedge test under a stricter locked holdout.
- Under DevelopmentPython· since 2026
Statistical Power Analysis of Single-Instrument Intraday Option Strategies
A working draft arguing that the in-sample-works, out-of-sample-fails pattern so common in retail options research is a sampling artifact of the sample sizes available, not evidence about any particular strategy.
- Required-sample-size calculation (Lo 2002) puts a concrete number on why single-instrument retail research can't reach significance
- Decomposes a 'profitable year' into its actual session-level distribution rather than reporting the total
- Reports two conflicting tests on the same data side by side instead of hiding the one that disagrees
- Documents and corrects its own earlier placebo-test error in the text
- Statistical Power
- Backtest Methodology
- Self-Correction
Not yet public. Delivered as a standalone manuscript (docx/pdf); reproduction scripts exist but haven't been packaged for release.
- CompletedPython· since 2026
India VIX and Next-Session Realized Range
Does today's India VIX predict tomorrow's realised range? Across 1,402 sessions it does (r = 0.469) — but a five-day-lagged VIX does nearly as well, so most of the effect is a slow regime read rather than a fresh daily signal.
- 1,402 sessions: r = 0.469, R² = 0.22
- Survives HAC standard errors, a block bootstrap and a locked holdout
- A five-day-lagged VIX still scores r = 0.389 — mostly a slow regime read, not a fresh daily signal
- Raw data committed: reproducible without a broker account
- India VIX
- Realized Volatility
- Newey-West Inference
- CompletedPython· since 2026
Black-Scholes Greeks and Implied-Volatility Attribution Engine
A from-scratch options pricer and Newton-Raphson implied-volatility solver, reconciled against realized P&L across more than 800 real trading sessions on two instruments.
- Black-Scholes pricing and Newton-Raphson IV solving built from first principles, not a wrapped library call
- Reconciled against realized P&L across 800+ real sessions on two instruments
- Per-leg, per-minute Greeks decomposition, not a single end-of-day number
- Used internally to attribute realized P&L to delta, gamma, theta and vega across the options-research pipeline
- Black-Scholes
- Greeks
- Implied Volatility
Internal infrastructure component, used across the options-research pipeline rather than published standalone.
- Active ResearchPython· since 2026
Overnight vs. Intraday Return Decomposition Across 184 NSE Stocks
A cross-sectional manuscript decomposing stock returns into overnight and intraday components across a 184-stock NSE panel, now in its tenth revision and moving toward SSRN.
- 184-stock NSE cross-sectional panel — the program's one deliberate move beyond single-instrument research
- Overnight and intraday return components decomposed and analyzed separately rather than as a single combined return
- Ten manuscript revisions to date; analysis pipeline migrated from Colab to a local environment for reliability
- Targeting SSRN publication alongside the existing OI-repositioning and power-analysis papers
- Cross-Sectional Research
- Overnight Returns
- NSE Equities
In progress. Local Claude Code pipeline; not yet packaged or pushed to a public repository.
- CompletedPython· since 2026
Donchian Breakout Option Overlay — A Falsification Study
An adversarial, holdout-tested study of the Donchian breakout as an option-buying signal, priced against 54,275 real NIFTY option legs, that finds no directional edge but confirms channel width forecasts move size out of sample.
- Adversarial, pre-registered design: signal built before choosing to buy or sell premium; 2024 fixes every structural choice, 2025 is in-sample, Jan–Apr 2026 is a locked holdout
- No directional edge at any horizon, in any of three periods, across 570 NIFTY spot sessions and a 93-cell conditioning grid indistinguishable from chance
- Every long-premium construction (straddle, directional single leg, short-strangle counterpoint) loses money in-sample and out of sample against 54,275 real 1-minute option legs
- Channel width — not direction — forecasts next-move size and holds out of sample, though it does not translate into buying premium being profitable
- Mirror short-premium iron condor test fully reverses sign in the holdout; slippage and forward-leak ruled out (see also: Transaction-Cost-Aware Condor Hedge)
- Donchian Channel
- Out-of-Sample Validation
- Options Hedging
- Negative Results
- Completed
intrabar
Open-source tooling that decides what filled, and at what price, inside an OHLC bar for bar-based backtests.
- Decides whether an order filled, and at what price, inside an OHLC bar
- Released publicly from the Equity Research Platform's fill engine
- Open Source
- Backtesting
- Execution Modelling
- Under Development
Equity Research Platform
A backtesting platform for NSE F&O stocks, built around a validated fill engine, a cost model calibrated to real contract notes, and an out-of-sample mode.
- Fill engine validated against an independent oracle
- Cost model calibrated to real contract notes
- Out-of-sample mode
- Fill engine released publicly as intrabar
- Transaction Costs
- Execution Modelling
- Backtesting
The platform is private. The repository link is its fill engine, released publicly as intrabar.
Repository· github.com(opens in a new tab) - Active Research
Event Pool and Event Atlas
A point-in-time database of about 127,000 NSE corporate and market events, with a development-period atlas of which event types could plausibly support a strategy.
- ~127,000 NSE corporate and market events, 2020–2026
- Point-in-time, drawn from 10 sources
- Development-period-only feasibility atlas
- Hard lock: no returns after 2022 viewed during mapping
- Event Studies
- Point-in-Time Data
Research directions
Stated intentions, not completed work. They appear here so the programme's direction is legible, and they will move above once something exists to show.
- Research Direction
Real-Time Market Analytics
Research direction: moving event detection and positioning analytics from batch reconstruction to live observation.
- Streaming ingestion of index and options-chain data
- Online structure detection with explicit confirmation lag
- Measurement of how much a live signal differs from its reconstructed counterpart
- Real-Time Systems
- Research DirectionPython· since 2026
Reinforcement-Learning Position-Management Study
An early-stage project asking why some discretionary NIFTY option scalpers stay net profitable after costs despite having no statistical edge on entries — by isolating trade management from entry choice using a learned policy.
- Deliberately isolates trade management from entry selection by testing a learned policy against random entries
- Targets a specific, testable explanation for observed discretionary trader profitability rather than a general RL exploration
- Built to reuse the existing 1-minute option dataset rather than requiring new data collection
- Currently a written environment specification, not yet implemented or tested
- Reinforcement Learning
- Position Management
- Behavioural Finance
Specification only (rl_trade_management_env_spec.md). Environment and policy not yet implemented.
Retired on evidence
Work that was built, tested and withdrawn. It stays visible because a method that failed is a finding, and the failures are the part of a research record that usually goes unpublished.
- ArchivedPython· since 2025
Day-Type Classifier — Retired on Evidence
A framework that reported a large gain, traced to look-ahead in its own gating rule, rebuilt three ways and retired when none of them survived out of sample.
- Full-session statistic used to gate an intraday decision — look-ahead in the gating rule itself
- Three causal reconstructions specified before testing, not after the failure
- All three profitable in sample, all three unprofitable out of sample across 264 sessions
- Retired rather than reported
- Out-of-Sample Validation
- Look-Ahead Bias
- Negative Results
Not published. The finding is the negative result; the framework it concerns was withdrawn from use.
- ArchivedPython· since 2026
Strike-Pinning and Price-Magnet Behavior Across NSE F&O Stocks
Tested whether NSE F&O stocks pin to their highest open-interest strike at monthly expiry. They do not: the close landed nearest the max-OI strike on about 10.1% of expiry days, against about 11.6% on other days.
- Monthly expiries, strike-level open interest and a five-year 1-minute price panel
- Close nearest the max-OI strike: ~10.1% of expiry days vs ~11.6% of other days (development, 2020–22)
- No pinning found — retired on evidence
- Method note: split/bonus-adjusted prices must be converted before comparing with raw bhavcopy strikes
- Cross-Sectional Research
- Strike Pinning
- Expiry Effects
- Open Interest
- ArchivedPython· since 2026
High-Trailing-Volatility Regime Gate — Retired on Evidence
The only conditioning rule in the condor program to survive a full adversarial battery, and still not a usable rule — it avoids losses rather than earning returns, fires unpredictably, and its calibration doesn't transfer across instruments.
- Passed rotation placebo, live-parameter walk-forward, 3x cost stress, and deflated Sharpe — then failed the quarterly walk-forward specifically designed to test it
- Three-quarters to four-fifths of its apparent value comes from avoiding losses, not generating returns
- Produced five consecutive years of zero trades when applied to BANKNIFTY
- The two instruments require opposite volatility-window constructions to work at all — a sign-dependent-on-estimator result
- Negative Results
- Regime Detection
- Walk-Forward Validation
Not published. The finding is retained as a negative result; the gate itself was withdrawn from use.
- ArchivedPython· since 2026
Directional Scalping Trigger Battery — Retired on Evidence
A wide sweep of classic directional intraday triggers — Alligator, MACD, ORB retest, TWAP cross, Bill Williams patterns, Golden Crossover, and more — tested against real 1-minute data and rejected almost without exception.
- Seven-plus classic directional triggers tested against real 1-minute data, all rejected as showing no edge over baseline
- Included independent re-verification of a trigger sourced from an external paid course, rather than accepting its claimed results
- Identified a structural data limitation (top-five-depth-only market data) ruling out true microstructure scalping at retail
- Directly produced the one surviving finding in the program — Donchian width as a sizing signal — by separating direction from magnitude
- Market Microstructure
- Negative Results
- Technical Indicators
Not published. Negative results retained internally; no structures from this batch were carried forward as direction signals.