Point-in-Time NSE Research · 2000–2026

Where Fancy AI Meets
Indian Market Reality.

Most trading AI backtests look miraculous until real NSE delivery costs, flat ₹15.34 depository charges, FIFO taxes, and delisted stocks are accounted for. We built a 26-year point-in-time simulator and pure-JAX engine to test 30+ hypothesis families—and show what actually survives.

Simulate Your Hurdle Rate ↓ Take the 6-Study Quiz ↓
8.77M
PIT Security-Session Rows
500K/s
Pure-JAX XLA Steps/Sec
3,865
Pre-Registered Trials
NSE Empirical Verdicts

Out-of-sample verdicts, net of statutory fees and ₹15.34 DP charges (select a study to inspect):

Post-Filing Earnings Drift (17 yrs) +5–7%/yr · Survives
Tax-Gated Hysteresis Band (20.7 yrs) +0.45%/yr · -58% Churn
LLM Filing Reader (1,563 filings) +1.4 pp · Below Cost Bar
High-Turnover Continuous RL (144 cells) -1.96 IR · Churn Wall
POST-FILING DRIFT · t > 2.5

Entered only after the filing was public. Stayed positive under measured spreads and doubled costs, above a high-lakh capital floor.

Test in Quiz ↓

Illustration: the simulator's 200-stock quarterly basket with an assumed 5%/yr gross edge, after public NSE fees, ₹15.34 DP and spreads:

-5.6%/yr
@ ₹1 Lakh
≈ ₹3L
Break-even
+3.2%/yr
@ ₹2 Crore

Every study runs on our deterministic point-in-time data plane and pure-JAX XLA stack:

1. PIT Data Plane
8.77M Rows
6,632+ archives (2000–2026), 1,142 delisted symbols (~30%), 24.2K corporate actions.
2. One-World Friction
1e-9 Pinned
Exact STT/Stamp/GST/₹15.34 DP waterfall and mandatory FIFO s.111A/s.112A tax lots.
3. Pure JAX + XLA
51× HBM Cut
8 RL algorithms, Patch-GTrXL sector transformer, 0.31 GB device FrameRingBuffer.
4. Anti-Overfit Gates
8× Lower Var
Paired active rescoring, SHA-256 frozen specs, and 3,865 Holm-Bonferroni trials.
Step 1 · Interactive NSE Cost & Tax Calculator

See Your True Retail Hurdle Rate

Pick a portfolio preset below (or drag any slider) to see how much gross return you need just to tie a passive NIFTY 500 Index Fund (~13.45%/yr after tax) once NSE delivery fees, the flat ₹15.34 DP sell charge, and FIFO capital-gains taxes are debited.

₹10 Lakh
₹1L₹2L₹5L₹10L₹20L₹50L₹1Cr₹2Cr
Quarterly (4×/yr)
Weekly (52×)Monthly (12×)Quarterly (4×)Annual (1×)
30 Stocks
10203050100200
4. Portfolio Construction Discipline
Gross Hurdle to Tie Index
17.9%/yr
Required gross to tie 13.45%/yr net index fund
HIGH FRICTION HURDLE
-4.45%/yr lost to fees & tax
Passive Index 13.45% Statutory + DP 0.71% Bid-Ask Spread 0.73% STCG/LTCG Tax 3.02%
Pick a preset or adjust the sliders to see why weekly/monthly AI trading and over-diversified small accounts hit an insurmountable mathematical cost wall on NSE.
Inspect Per-Trade Fee & Tax Breakdown ₹33,333/stock · 55.2 bps RT
Position Ticket Size
₹33,333
₹15.34 DP fee = +4.6 bps
All-In Round-Trip Cost
55.2 bps
27.2 bps statutory + 28.0 bps spread
Annual Turnover & Tax Realization Drag
2.6×/yr turnover + 3.02% tax wedge

Next: predict 6 real studies ↓

Step 2 · Interactive 6-Study Benchmark Quiz

Test Your Quant Intuition (One Study at a Time)

Popular financial media and AI papers make bold claims. We tested them on survivorship-free Indian market history with full statutory costs. Make your prediction on Study 1 below, then step through at your own pace.

Study 1 of 6 · 0 / 6 Answered
Study 01 of 06 · AI Text vs. Accounting Numbers Question 1 / 6

Does an AI model reading 1,563 corporate filings predict next-quarter earnings direction better than a numbers-only baseline?

↑ Pick one option to reveal the empirical NSE result

Almost a tie: 74.2% vs. 72.8% (+1.4 pp)

Across 1,563 exchange filings (with company names, dates, and stock prices masked to prevent LLM memorization), reading prose added just +1.4 percentage points over past quarterly numbers alone—far below our 5.0 pp pre-registered hurdle needed to survive 53+ bps round-trip trading costs.

Study 02 of 06 · Insider Activity (66 Months) Question 2 / 6

Does buying NSE stocks after open-market promoter purchases (held 60 sessions) beat a matched industry & liquidity benchmark?

↑ Pick one option to reveal the empirical NSE result

Fails: -2.21%/yr net (t = -0.61)

Over 66 months (2020–2026), entering after public disclosure of open-market promoter buys trailed matched controls by -2.21%/yr net (-3.43%/yr at ₹5k ticket costs). The 95% upper confidence bound (+3.84%/yr) formally refuted published literature claims of +16%/yr.

Study 03 of 06 · 25,166 Large Daily Moves Question 3 / 6

When a top-300 NSE stock jumps ≥5% in a single day, is there usually an exchange news filing—and does buying the spike continue?

↑ Pick one option to reveal the empirical NSE result

72.4% of ≥5% moves have zero filing; intraday spikes reverse -57.2 bps

Across 25,166 ≥5% stock-days, 72.4% had no exchange filing at all (and only 20.4% had an informative filing). Chasing 3–5% intraday filing spikes at a realistic 5-minute retail execution delay lost -57.2 bps by the close. Across 893 tested daily/intraday event cells, 892 were underpowered after 86.7 bps all-in costs.

Study 04 of 06 · Continuous RL Positive Control (144 Cells) Question 4 / 6

If a continuous-action RL agent is fed a calibrated next-bar signal on daily NSE bars, does it convert that signal into net after-cost profit?

↑ Pick one option to reveal the empirical NSE result

Fails: -1.96 OOS IR (~710 trades/episode)

Across 144 pre-registered cells (12 seeds × 6 walk-forward folds), standard continuous-action RL policies churned ~710 trades per evaluation window—converting signal into statutory fee drag (-1.96 IR vs -2.04 sham; paired difference +0.076 IR, t = 2.31 < 2.57 critical). Without hard turnover hysteresis, high-frequency rebalancing drowns in NSE delivery friction.

Study 05 of 06 · Low-Turnover Fundamental Drift (208 Months) Question 5 / 6

Does a low-turnover earnings-surprise strategy, entered only after the filing is public, survive 2× costs and live spreads?

↑ Pick one option to reveal the empirical NSE result

Works: +5–7%/yr net (t > 2.5), still positive after live spreads

Over 208 months (2009–2026), entered only after the filing was public, it beat matched controls by +5–7%/yr net (significant after Holm multiple-testing correction) and stayed positive under both measured live spreads and an adversarial 2× doubled-cost audit—but only for accounts above a high-lakh capital floor.

Study 06 of 06 · Direct 300-Stock Basket vs. Index Fund (20.7 Years) Question 6 / 6

Does managing your own equal-weight 300-stock NSE equity basket beat buying a single low-cost NIFTY 500 Index Fund after costs and FIFO taxes?

↑ Pick one option to reveal the empirical NSE result

Index Fund wins in 69 of 70 (Capital × Breadth) tiers

Over 20.73 years (2006–2026), a self-held 300-stock basket suffered 2.14%/yr of cumulative rebalancing costs and annual FIFO capital-gains tax realizations (10.48%–10.93%/yr net), trailing a passive NIFTY 500 Index Fund (11.67%/yr net after 0.42% tracking drag and a single terminal LTCG tax redemption) across nearly every capital tier from ₹2 Lakh to ₹2 Crore.

Prefer to skim?
Step 3 · The Empirical Ledger · 3,865 Trials · 30+ Hypothesis Families

What Survived vs. What We Closed (And Why)

In quantitative research, knowing what does not work after costs is just as valuable as knowing what does. Start with the 3 surviving architectures below, then pick a category to inspect what failed and why.

We publish what we tested, how we tested it, and whether it survived — not how to trade the survivors.

3 Survived All Gates — Low-turnover post-filing drift & tax-gated hysteresis 11 Cost & Tax Wall — Gross signal existed, net return negative after NSE fees/FIFO tax 14 Zero OOS Edge — Indistinguishable from turnover-matched null controls 6 Timing / Universe Artifact — Edge vanished once survivorship or broadcast timestamps were fixed
SURVIVED · t > 2.5 (2009–2026)

Post-Filing Earnings Surprise Sleeve

Beat matched controls by +5–7%/yr net of statutory fees and DP charges, and stayed positive under measured spreads and 2× costs.

Entry · Selection · Holding — withheld

Inspect Study 05 in Quiz ↑
SURVIVED · -58% Turnover

Hysteresis Band + LTCG Tax-Gate Architecture

Combining a multi-threshold retention buffer, no-trade rebalance bands, cash buffer, and idle-cash sweep cut trade count by 58% and recovered +39 to +45 bps/yr of after-tax compounding.

Band widths · Thresholds — withheld

Simulate Tax-Gated Mode ↑
SURVIVED · +1.52%/yr vs Matched Null

Turnover-Matched Low-Volatility Tilt

Cleared a 20-seed persistence- and turnover-matched random control by +1.52%/yr on the survivorship-free PIT panel when rebalanced conservatively.

Universe · Rebalance rule — withheld

Read Control Methodology →
ℹ

We are a research-stage startup. No real capital is deployed. As of October 2026, no method — reinforcement learning included — has cleared our own evidence bar for live trading, and we say so. 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.

Collaborators can request full study details under NDA →

Institutional Infrastructure · Built From Scratch

Why Our Verdicts Hold Up: The Engineering Moat

You cannot measure a 3%–6%/yr market edge inside a simulator that leaks 5%/yr of survivorship bias or ignores statutory tax lots. Scan the 6 pillars below, or expand any card for its full technical specification.

8.77M-Row Point-in-Time (PIT) Panel
Reconstructed 26 years of NSE history (2000–2026) including 1,142 delisted companies (~30%) that standard broker APIs silently drop.
Technical Specification

Standard broker feeds omit 72% of active 2018 symbols, inflating naive backtests by +2.31% to +6.00%/yr. We rebuilt the entire panel from 6,632+ raw exchange archives across 4,277 continuous security chains, 4,561 ISIN mappings, and 24,203 corporate actions.

6,632+ ARCHIVES · 1,142 DELISTED SYMBOLS
"One-World" Statutory & FIFO Tax Engine
Identical rupee-exact fee and FIFO demat tax-lot accounting shared across backtest, paper, and live execution—pinned to 1e-9 precision.
Technical Specification

Verified by 23 golden test suites: dated Finance Act STT schedules, stamp duty, 18% GST, flat ₹15.34 CDSL DP charges, and bitemporal Indian FIFO demat tax lots (20.8% STCG vs 13.0% LTCG with 1 Feb 2018 grandfathering).

23 GOLDEN PIN SUITES · FIFO s.45(2A)
Pure-JAX + Optax XLA Vectorized Training
8 RL algorithms compiled via jax.jit, jax.vmap, and fused jax.lax.scan kernels for 500,000 env-steps/sec stepping.
Technical Specification

Zero PyTorch overhead across SAC, Temporal SAC, TQC, TD3, PPO, A2C, DQN, QR-DQN, and IQN—delivering ~38,300 env-steps/sec full 128-env GPU training throughput (4.49× scan speedup with zero recompiles).

500K ENV-STEPS/SEC · ZERO RECOMPILES
Patch-GTrXL, Ragged-Dot MoE & 51× HBM Cut
15-sector NSE graph attention transformer + custom FrameRingBuffer that compresses GPU replay memory by 51× (314 GB → 0.31 GB).
Technical Specification

Combines cross-asset temporal attention with jax.lax.ragged_dot Top-K Mixture-of-Experts (up to 48.6M parameters) while reconstructing 32-frame temporal stacks on-the-fly in device HBM.

48.6M PARAMS · 51× REPLAY COMPRESSION
SHA-256 Pre-Registration & 8× Variance Cut
Every confirmatory study is frozen via SHA-256 before evaluation and scored against paired identical-window controls across 3,865 logged trials.
Technical Specification

Enforces an 80%-power Minimum Detectable Effect (MDE80) gate before outcomes are read, achieving an 8× cross-fold variance reduction under Holm-Bonferroni familywise error rate control.

SHA-256 FROZEN SPECS · 3,865 LOGGED TRIALS
6-Layer Pre-Trade Gate & AST Verification
Protected by 37,000+ automated tests, 44 AST lookahead gates, and a < 5ms pre-trade risk kill switch.
Technical Specification

Static AST checks block future-timestamp access at commit time, while the 6-layer execution gate enforces circuit & fat-finger price collars and SEBI retail-algo rate limits (≤ 10 orders/sec).

37,000+ TESTS · 44 AST GATES · <5ms KILL SWITCH
AI-Native Engineering

Researched and Built With AI Agents.

One founder cannot hand-write and hand-check a trading platform of this size. We work with AI research and coding agents — several models, not one vendor — under written rules, with a human decision at every gate.

Many Models, One Rulebook

We pick the model that fits the job: hosted models such as Claude, Codex, Gemini and Kimi for research, planning and code, and open-weight models such as GLM and Qwen on our own GPU for bulk document work. Every agent follows the same written rules, and the founder decides at every gate that involves risk or money.

Multi-model Human at every gate

Agents That Check Each Other

A lead agent splits work into written briefs. Builder agents implement them. A separate reviewer agent, which did not write the code, must accept the change before it merges.

Multi-agent workflow Independent review

Proof, Not Claims

Every fix ships a test that fails before the fix and passes after it. Reviewers plant deliberate defects to confirm that the tests can catch them.

Fail-before tests Planted-defect checks

Designs Frozen Before Data

Each study's design and code are written down and hash-frozen before the first data read, so a result cannot be tuned after the fact.

Pre-registered studies Hash-frozen designs
ℹ

No AI model places trades. All strategies run in simulation and paper trading, with no live capital. In progress: extracting structured data from company filings and exchange announcements with document models. Model and product names belong to their owners; RL Alpha Labs is an independent user of these tools, not a partner or affiliate of any of them.

Read our AI agent lab rules →

Forensic Validation · Why Most Retail Backtests Lie

The 8 Backtest Traps We Caught and Eliminated

Each trap below inflated simulated Indian equity returns by +0.3%/yr to +29.6%/yr before our invariant audits caught and sealed it. Click any row to inspect the failure mode and our engine rule.

01 Broker Store Survivorship Bias +2.3% to +6.0%/yr phantom alpha

Standard broker APIs drop delisted tokens. Selecting a universe from today's active list silently erases 1,142 dead NSE companies (like Satyam, DHFL, or Future Retail) and inflates annual returns by up to 6 percentage points.

Engine RuleEvery historical rebalance queries our 6,632-archive point-in-time panel strictly as-of session t, pricing terminal delistings explicitly.

02 Same-Bar Close Decision & Fill +8.0% to +29.6%/yr phantom alpha

Computing a signal from day t's closing price and filling the trade at that exact same close captures a phantom microstructure bounce that no live order can ever achieve.

Engine RuleMandatory 1-session execution lag: signals computed after close t execute at session t+1 with upper/lower circuit-band lock checks.

03 Quarter-End vs. Broadcast Timestamp Leakage +0.75%/yr phantom alpha

Assuming Q4 financials are known on March 31 (or on the statutory deadline) leaks future earnings weeks before the company actually broadcasts its XBRL filing to the exchange.

Engine RuleEvery fundamental record is gated by knowable_at = max(received_ts, broadcast_ts) mapped to the next NSE trading session after 15:30 IST.

04 Ignoring the Flat ₹15.34 CDSL DP Charge +30.7 bps/sell @ ₹5k ticket

Modelling delivery costs as a flat percentage (e.g., 10–20 bps) hides the ₹15.34 per-scrip depository sell fee—which adds +30.7 bps on a ₹5,000 position and destroys small-capital multi-stock portfolios.

Engine RuleFull per-fill statutory waterfall with exact ₹15.34 DP debit on delivery sells and NSE dated tick-size floors.

05 Illegal Tax-Lot Selection & Pre-Tax Comparisons +0.30% to +3.0%/yr tax omission

Comparing a high-turnover strategy pre-tax against an index, or using Highest-Cost-First (HCF) lot accounting, hides the 20.8% STCG tax paid every rebalance. Indian law (s.45(2A)) mandates FIFO per demat account.

Engine RuleBitemporal FIFO demat tax-lot accounting with historical Finance Act rates, 8-year loss carry-forward, and terminal liquidation tax.

06 High-Churn Random Baselines +1.0% to +1.4%/yr false flattery

Comparing a low-turnover factor portfolio against a naive random-stock baseline that replaces 100% of its holdings every quarter makes the candidate look good solely because the random baseline burned 1.4%/yr in extra fees.

Engine RuleAll null controls are persistence-, turnover-, and liquidity-tercile matched to the candidate strategy.

07 Overlapping-Window Sign-Share Illusions 50.15% false-positive rate

Counting how many rolling 3-year or 5-year windows are positive ("beat the market in 73% of rolling windows") treats 95%-shared data as independent—producing a 50.15% false-positive rate under zero true skill.

Engine RuleDisjoint fold evaluation, Newey-West HAC standard errors, and noncentral-t 80%-power MDE gates.

08 LLM Entity & Price Memorization Lookahead via pre-training

Feeding raw historical news or filings to a modern LLM allows it to recognize the company name, year, and stock price from its pre-training corpus—leaking future returns disguised as "sentiment analysis."

Engine RuleAutomated entity, date, and price scrubber replaces all tickers/ISINs with [FILER], dates with [D-30], and strips market-quote clauses before inference.

Technical Documentation · Zero Stubs

Explore the Full Research & Engineering Papers

Dive into our mathematical formulations, NSE market microstructure tables, pure-JAX training benchmarks, and curated literature review.

01 · Problem Formulation & NSE Microstructure
Formal POMDP/MDP definition, 2015–2025 Indian market regime breakdown, NIFTY 50/100/500 liquidity & spread tables, and statutory CNC vs. MIS transaction cost equations.
READ TECHNICAL PAPER →
02 · Three-Plane Architecture & Methodology
How our Data Plane, Pure-JAX Model Plane, and One-World Simulation Plane fit together—plus multi-component risk-aware reward shaping and stage-gated promotion rules.
READ ARCHITECTURE SPEC →
03 · JAX XLA Scaled Training & Throughput Benchmarks
Hardware benchmarks across vectorized environments, fused jax.lax.scan loops, Patch-GTrXL sector-graph transformers, and 51× replay buffer compression.
VIEW BENCHMARK RESULTS →
04 · 20 Open Research Questions
Our structured catalog of 20 open problems across Fundamental Signal Limits, Low-Turnover Policy Design, Bitemporal Feature Engineering, and Tax-Aware Portfolio Control.
EXPLORE 20 OPEN PROBLEMS →
05 · Annotated Literature & NSE Commentary
17 foundational papers in RL for portfolio management, meta-learning, market microstructure, and distributional risk—annotated with what holds and what breaks on Indian equities.
READ ANNOTATED BIBLIOGRAPHY →
06 · Field Notes: Indian Trading Costs & Backtest Hygiene
Plain-language deep dives into Indian statutory fee math, the 8 backtest traps, closed hypothesis post-mortems, and deterministic verification gates.
READ FIELD NOTES →
The Lab

Independent, Self-Funded & Evidence-First

Rahul
Rahul
Founder · Systems & Quantitative Research
India · NSE / BSE Equities

RL Alpha Labs is an independent quantitative research and systems engineering lab focused on Indian equities. We do not sell courses, trading tips, or black-box retail subscriptions. Every claim on this site is backed by reproducible code, point-in-time exchange archives, and SHA-256-pinned evaluation ledgers.

LinkedIn ↗ GitHub ↗ Research Papers → Join the lab →
8.77M Rows
Point-in-time NSE daily security-session observations (2000–2026) with ~30% delisted symbol coverage.
100% Pure JAX
XLA-compiled vectorized environments, Patch-GTrXL sector-graph transformers, and 51× compressed device ring buffers.
37,000+ Tests
Automated unit, contract, property, golden-pin, and AST architecture gates enforcing zero lookahead leakage.
3,865 Trials
Every scored configuration logged in an append-only ledger under Holm-Bonferroni familywise error rate control.
Get In Touch

Collaborate on Research, Data, or Systems

Whether you are a quantitative researcher working on Indian market microstructure, a systems engineer interested in high-throughput JAX/XLA simulation, or an institutional partner exploring licensing our certified PIT data and verification stack, we would love to hear from you.

Research Collaboration
Joint empirical studies on our 26-year survivorship-free NSE harness
Institutional & B2B Licensing
Certified PIT datasets, statutory cost/tax engines & verification tooling
Systems & ML Engineering
JAX/XLA acceleration, custom transformers & pre-trade risk architecture

Opens your email app with this message pre-filled — nothing is sent until you press send there. Prefer webmail? Write to connect@rlalphalabs.com or reach out on LinkedIn ↗.

✓

Message Prepared

Your email app should have opened with your message — press send there to deliver it. Nothing opened? Write to connect@rlalphalabs.com directly, or connect on LinkedIn ↗.