Nifty 500 weekly ranking with a pooled LightGBM: build, optimisation and test, 2010 to September 2025
Update 31 Oct 2026 (version 2). Frozen v2 re-run on survivorship-free data: validation 2018 to September 2025 21.3% at 1x and 18.7% at 2x (earlier 25.6%), development 2010 to 2017 10.7% (earlier 16.0%) with a 50% drawdown in 2011; correlation with the two Longbow strategies about 0.45. Read the survivorship-free re-test, comparison and Kelly analysis. The figures below are the earlier version (Dhan price history without delisted stocks, 4%/yr allowance) and are kept for the record.
Short answer. The pooled model has real, persistent ranking skill: over 16 out-of-sample years the top predicted decile beat the bottom one by 0.76% a week, monotonically and in every year, and the strategy built on it beat random picks from the same stock universe by a wide margin in both periods. Two portfolio versions were frozen on 2010-2017 data and run once on 2018 to September 2025. Both pass every pre-registered gate; the lower-turnover v2 (N = 30, rebalance Friday every second week, buffer k = 2.0, liquidity floor ₹5 crore, 5-session model) returned 25.6% a year net against 11.8% for the Nifty 500 price index (still 15.9% after the 4% survivorship allowance at three times the assumed costs), with a similar drawdown (36.2% vs 34.4%). The development margin was smaller (16.0% against 10.1%; alpha t = 0.8 after the allowance), so the validation margin should not simply be extrapolated. Three things keep this a candidate, not a proven edge: most of the return comes from small and mid-sized, less liquid stocks in a period when they boomed (2020 to 2023); 6 to 26% of past index members have no price history in our data (survivorship); and the weekly index-direction call has no skill (accuracy 54% against a 59% always-up baseline). The sealed period (October 2025 to October 2026) was opened once, at your instruction: in a falling market (index -5.2%) v2 gained 4.5% a year net against random picks' -8.0%; a forward paper run of v2 starts on 2 November 2026.
1. What the write-up asks for, what we kept and what we changed
| Write-up | Here | How |
|---|---|---|
| Pooled LightGBM, 5-day forward return, cross-sectional ranks, market context, VIX interactions | Kept | 31 features: 17 within-day ranks, 10 market-context features, 4 rank × VIX-regime interactions |
| Purged expanding walk-forward, early stopping, regularised shallow trees | Kept, with 16 yearly folds (2010-2025) instead of 4 | Embargo 6 sessions (11 for the 10-session model); three seeds averaged |
| Label from the rebalance close | Changed | Label is the return from the next open to the open five sessions later, matching how a close-based signal can actually be traded |
| Target: raw return | Changed | Target is the cross-sectional rank of that return (market component removed); market direction is a separate head |
| Holding buffer, N sweep, weekday sweep, empty slots in cash | Kept and extended | Buffer k up to 3, liquidity floor, 10-session model blend, weekly or fortnightly cadence (version 2) |
| Cost: square-root impact | Changed | Indian delivery charges plus slippage by liquidity tier (v1: 0.05/0.10/0.20%; v2: 0.05/0.10/0.25/0.50%), stressed at 2x and 3x |
| Selection by robustness and low turnover | Kept, with two safeguards | Selection among configurations positive in at least 60% of development years (and, in v2, alpha t above 1.5 at 2x costs and at least 4 turnovers a year) |
| Index-direction call from breadth of predictions | Changed | Separate small model on market features; evaluated honestly (section 8) |
2. Data and test period
- Period: predictions are out-of-sample for 2010-01 to 2025-09, 15.75 years; training starts in 2004. Earlier work on these data tested 2018 to 2025 only. Development = 2010-2017 (all 540 configurations run once; selection), validation = 2018 to September 2025 (one run per frozen version), sealed = October 2025 onward (not loaded).
- Universe: Nifty 500 members on each date (point-in-time from 1998), 20-day median traded value at least ₹0.5 crore for the model, and the version's own floor for the portfolio. Daily prices from Dhan, split and bonus jumps removed by the project's rule (a one-day move above 35% is not counted, which also removes real crashes of that size).
- Survivorship, measured: the share of Nifty 500 members that have price history in the panel was 60% in 2010, 71% in 2017, 79% in 2020 and 94% in 2025. Missing names are delisted or merged companies, which would have pulled returns down. A 4%/yr allowance is deducted in the gates; results with a larger allowance are in section 9.
| Year-end | Index members | With price history | Coverage |
|---|---|---|---|
| 2010 | 500 | 298 | 60% |
| 2011 | 500 | 306 | 61% |
| 2012 | 500 | 315 | 63% |
| 2013 | 500 | 321 | 64% |
| 2014 | 500 | 325 | 65% |
| 2015 | 500 | 334 | 67% |
| 2016 | 501 | 346 | 69% |
| 2017 | 501 | 357 | 71% |
| 2018 | 501 | 369 | 74% |
| 2019 | 501 | 388 | 77% |
| 2020 | 501 | 395 | 79% |
| 2021 | 501 | 417 | 83% |
| 2022 | 501 | 427 | 85% |
| 2023 | 501 | 440 | 88% |
| 2024 | 501 | 454 | 91% |
| 2025 | 501 | 469 | 94% |
3. The ranking model
Average daily rank IC across the 16 test years: 0.076 (range 0.041 to 0.114). Top-decile minus bottom-decile spread: 0.76% per five sessions overall. In stocks with more than ₹100 crore of daily traded value the top decile earns about the same as the average stock (0.45% vs 0.32%); the long-side edge sits in the ₹5 to ₹100 crore range and below. This is why the liquidity floor and the cost tiers matter.
4. Portfolio versions and results
v1 (as inspired, v1 costs): N = 30, rebalance Friday every week, buffer k = 1.0, liquidity floor ₹0.5 crore, 5-session model. v2 (improved, v2 costs): N = 30, rebalance Friday every second week, buffer k = 2.0, liquidity floor ₹5 crore, 5-session model. Both configurations were chosen on 2010-2017 only, by the pre-registered rules; validation was run once for each, together.
v2: the recommended version
| Period | Costs | Net CAGR | After 4%/yr allowance | Nifty 500 price | Sharpe | Max drawdown | Index drawdown | Alpha after allowance | Years ahead of index | Turnover (× capital a year) | Costs a year |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Development 2010-2017-12 | 1x | 16.0% | 12.0% | 10.1% | 1.00 | 25.8% | 30.4% | 1.8% (t 0.8) | 75% | 4.1 | 1.3% |
| Development 2010-2017-12 | 2x | 14.3% | 10.3% | 10.1% | 0.91 | 27.8% | 30.4% | 0.4% (t 0.2) | 62% | 4.2 | 2.7% |
| Development 2010-2017-12 | 3x | 12.6% | 8.6% | 10.1% | 0.82 | 29.7% | 30.4% | -1.0% (t -0.5) | 62% | 4.2 | 4.0% |
| Validation 2018-2025-09 | 1x | 25.6% | 21.6% | 11.8% | 1.22 | 36.2% | 34.4% | 7.1% (t 2.2) | 88% | 7.4 | 2.0% |
| Validation 2018-2025-09 | 2x | 22.7% | 18.7% | 11.8% | 1.12 | 36.0% | 34.4% | 4.9% (t 1.5) | 88% | 7.5 | 4.0% |
| Validation 2018-2025-09 | 3x | 19.9% | 15.9% | 11.8% | 1.00 | 37.0% | 34.4% | 2.6% (t 0.8) | 75% | 7.5 | 6.0% |
v1: close to the original design
| Period | Costs | Net CAGR | After 4%/yr allowance | Nifty 500 price | Sharpe | Max drawdown | Index drawdown | Alpha after allowance | Years ahead of index | Turnover (× capital a year) | Costs a year |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Development 2010-2017-12 | 1x | 22.0% | 18.0% | 10.1% | 1.37 | 21.9% | 30.4% | 7.6% (t 3.0) | 88% | 21.1 | 6.4% |
| Development 2010-2017-12 | 2x | 13.8% | 9.8% | 10.1% | 0.92 | 29.3% | 30.4% | 0.7% (t 0.3) | 88% | 21.5 | 13.1% |
| Development 2010-2017-12 | 3x | 6.2% | 2.2% | 10.1% | 0.46 | 42.0% | 30.4% | -6.3% (t -2.5) | 12% | 21.8 | 20.1% |
| Validation 2018-2025-09 | 1x | 29.0% | 25.0% | 11.8% | 1.26 | 39.7% | 34.4% | 9.7% (t 2.3) | 88% | 28.1 | 7.0% |
| Validation 2018-2025-09 | 2x | 19.2% | 15.2% | 11.8% | 0.91 | 41.3% | 34.4% | 1.8% (t 0.4) | 62% | 28.5 | 14.5% |
| Validation 2018-2025-09 | 3x | 10.1% | 6.1% | 11.8% | 0.55 | 49.8% | 34.4% | -6.1% (t -1.5) | 50% | 28.8 | 22.3% |
| Year | v2 net return | v1 net return | Nifty 500 price index |
|---|---|---|---|
| 2010 | 17.7% | 22.7% | 13.1% |
| 2011 | -18.4% | -14.1% | -27.2% |
| 2012 | 32.3% | 24.0% | 31.8% |
| 2013 | 16.2% | 20.7% | 3.6% |
| 2014 | 49.0% | 70.3% | 37.8% |
| 2015 | 10.3% | 9.3% | -0.7% |
| 2016 | 0.5% | 14.5% | 3.8% |
| 2017 | 33.5% | 45.7% | 35.9% |
| 2018 | 0.7% | 3.2% | -3.4% |
| 2019 | 7.6% | 3.5% | 7.7% |
| 2020 | 50.3% | 49.7% | 16.7% |
| 2021 | 60.5% | 76.5% | 30.2% |
| 2022 | 8.9% | 24.6% | 3.0% |
| 2023 | 46.2% | 53.3% | 25.8% |
| 2024 | 25.5% | 30.2% | 15.2% |
| 2025 | 12.0% | 2.2% | 1.6% |
2025 is January to September; 2018 to 2025 returns restart from cash on 1 January 2018.
5. Optimisation: what the grids showed
- Turnover is the lever. Without a buffer the book turns over about 70 times a year and costs eat most of the edge; k = 1 cuts costs to about 7% a year, k = 2 with fortnightly rebalancing to about 1.3% (v2).
- Liquidity floors trade return for robustness. Raising the floor from ₹0.5 to ₹5 to ₹20 crore lowers mean development CAGR from 21.9% to 17.3% to 13.3% (1x v2 costs) but makes the result less dependent on the slippage assumption.
- Weekday and N matter little once the buffer is wide (spread across weekdays and between N = 20 and 30 is within a few points).
- The 10-session blend and fortnightly cadence both help a little at 2x costs.
- Degenerate setting caught: with k = 3 the book is almost never refreshed (buy-and-hold of the first picks); the selection rule was corrected before validation to require at least four turnovers a year (disclosure 9.1 in the specification).
6. Controls
| Version | Random selection, development (median / 90th pct) | Random selection, validation | Model, development | Model, validation | 20-day momentum / 5-day reversal ranks, validation | With index-head cash overlay, validation | Shifted sample 2012-2025 vs index |
|---|---|---|---|---|---|---|---|
| v2 | 5.9% / 8.2% | 11.0% / 14.2% | 16.0% | 25.6% | 23.7% (dd 58%) / 7.9% | 18.7% (t 0.7) | 22.4% vs 14.1% |
| v1 | -5.6% / -3.1% | -1.7% / 0.6% | 22.0% | 29.0% | 11.9% (dd 77%) / -0.5% | 25.0% (t 1.8) | 29.4% vs 14.1% |
- Random selection (same universe, N, cadence, buffer and costs, 100 seeds) earns about the index level or less; the model is far above its 90th percentile in both periods. That is the strongest evidence of ranking skill.
- Simple baselines do not do the same job consistently: 20-day momentum ranking earned 23.7% in validation (with a 58% drawdown) but only 7.6% in development; 5-day reversal lost money after costs.
- Reality check across all 540 development configurations: best t = 5.29, p = 0.000. Plateau: every one-step neighbour of the frozen configurations was positive in both periods.
- Market regimes (validation): the strategy gained over the index in low and medium-VIX weeks. In the nine weeks of the highest VIX regime (March 2020) it fell with the market and slightly more than the index.
7. Pre-registered pass rule
| v2 condition | Result | Figures |
|---|---|---|
| 1. Validation net CAGR after the 4%/yr allowance at least the index, alpha t above 1.5 after the allowance | pass | 21.6% vs 11.8%; t = 2.16 |
| 2. Net excess over the index positive in at least 60% of validation years | pass | 88% of years |
| 3. Still ahead of the index (after the allowance) at 2x costs | pass | 18.7% vs 11.8% |
| 4. Development CAGR above the 90th percentile of random selection; reality check p below 0.10 | pass | 16.0% vs null p90 8.2%; reality-check p = 0.000 |
| 5. Maximum drawdown within the index's plus 10 points; worst year above −25% | pass | drawdown 36.2% vs index 34.4%; worst year 0.7% |
| 6. Plateau: at least 60% of neighbouring configurations positive in both periods | pass | 100% of 11 neighbours |
| 7. Shifted sample (first two years dropped) keeps the sign | pass | alpha 3.1%, t = 1.31 |
| v1 condition | Result | Figures |
|---|---|---|
| 1. Validation net CAGR after the 4%/yr allowance at least the index, alpha t above 1.5 after the allowance | pass | 25.0% vs 11.8%; t = 2.33 |
| 2. Net excess over the index positive in at least 60% of validation years | pass | 88% of years |
| 3. Still ahead of the index (after the allowance) at 2x costs | pass | 15.2% vs 11.8% |
| 4. Development CAGR above the 90th percentile of random selection; reality check p below 0.10 | pass | 22.0% vs null p90 -3.1%; reality-check p = 0.000 |
| 5. Maximum drawdown within the index's plus 10 points; worst year above −25% | pass | drawdown 39.7% vs index 34.4%; worst year 2.2% |
| 6. Plateau: at least 60% of neighbouring configurations positive in both periods | pass | 100% of 6 neighbours |
| 7. Shifted sample (first two years dropped) keeps the sign | pass | alpha 8.5%, t = 3.13 |
All seven conditions pass for both versions at their assumed costs. v1 fails at 3x costs (10.1% against the 11.8% index) while v2 still earns 19.9% before and 15.9% after the allowance: another reason to prefer v2.
7b. Sealed hold-out, opened once (October 2025 to 28 October 2026)
On your instruction the sealed period was opened one time on the two frozen configurations (ledger: research/rank500/holdout_ledger.json). One model was trained on rows whose label ended before 1 October 2025 (same parameters and seeds) and the strategies started from cash. No gate was written for this period; it is reported as the last independent check. Window length is about 13 months, so statistical power is low.
| Version | Net CAGR 1x costs | 2x | 3x | After 4%/yr allowance | Nifty 500 price index | Max drawdown vs index | Alpha after allowance | Random selection (median / 90th pct) | Turnover |
|---|---|---|---|---|---|---|---|---|---|
| v2 | 4.5% | 3.1% | 1.8% | 0.5% | -5.3% | 16.5% vs 13.1% | 7.6% (t 1.2) | -8.0% / -2.0% | 5.6 |
| v1 | 10.9% | 5.2% | -0.2% | 6.9% | -5.3% | 14.0% vs 13.1% | 12.8% (t 2.0) | -19.0% / -11.5% | 23.0 |
In a market that fell 5.2% over the window, v2 earned 4.5% a year net (the model beat the 90th percentile of random selection, -2.0%) but only 0.5% after the survivorship allowance, with alpha t = 1.2, below the 1.5 used in the validation gate. v1 earned 10.9% but fell to -0.2% at three times costs. The sealed period therefore supports the ranking skill (both versions are far above random picks and the index) without confirming a large absolute return; it also shows again that v2's low turnover is what keeps it cost-robust.
7c. Charts and trade details (v2, all three windows)
The same v2 book in each window: growth against the index, drawdown, how much of the book was invested, the shape of the trades, where the money was made, and the largest winners and losers. Per-trade, per-fill and per-day files are on the strategy page, which has the interactive equity chart, monthly table and the full trade log.
Which names made and lost the money
Across all three windows the 10 best names contributed 20% of total net P&L of ₹30,831,434 (394 names traded, 1594 round trips).
| Best 10 names | Trades | Net P&L | Avg trade |
|---|---|---|---|
| LUPIN | 10 | ₹842,592 | 31.3% |
| CHOICEIN | 2 | ₹833,453 | 74.0% |
| GLENMARK | 16 | ₹670,608 | 8.0% |
| EICHERMOT | 7 | ₹639,013 | 31.9% |
| TATAMOTORS | 7 | ₹573,477 | 21.0% |
| TRENT | 5 | ₹547,474 | 16.2% |
| MARUTI | 7 | ₹504,330 | 24.9% |
| TATAELXSI | 8 | ₹501,734 | 20.3% |
| BAJAJ-AUTO | 9 | ₹496,212 | 13.4% |
| MCX | 4 | ₹485,394 | 19.6% |
| Worst 10 names | Trades | Net P&L | Avg trade |
|---|---|---|---|
| HEROMOTOCO | 12 | ₹-316,378 | -3.1% |
| OFSS | 11 | ₹-224,226 | -6.3% |
| CHAMBLFERT | 7 | ₹-219,914 | -1.7% |
| KIMS | 2 | ₹-214,896 | -18.2% |
| INDIACEM | 6 | ₹-212,880 | -4.7% |
| SIEMENS | 7 | ₹-210,552 | -0.0% |
| UTIAMC | 1 | ₹-198,435 | -22.7% |
| NTPC | 7 | ₹-196,922 | -8.4% |
| BALRAMCHIN | 7 | ₹-196,777 | -2.0% |
| VTL | 1 | ₹-195,192 | -41.3% |
Ten best and ten worst round trips
| Name | Entry | Exit | Days | Entry price* | Exit price* | Return | Net P&L | Costs | Window |
|---|---|---|---|---|---|---|---|---|---|
| CHOICEIN | 2023-10-16 | 2024-11-25 | 406 | 2.83 | 7.29 | 156.2% | ₹876,566 | ₹5,621 | validation |
| MARUTI | 2014-01-20 | 2016-12-26 | 1071 | 4.73 | 13.81 | 191.2% | ₹462,418 | ₹1,913 | development |
| GLENMARK | 2025-02-17 | 2025-09-08 | 203 | 36.21 | 56.57 | 55.7% | ₹437,116 | ₹4,217 | validation |
| TATAMOTORS | 2023-11-13 | 2024-07-29 | 259 | 5.51 | 9.49 | 71.6% | ₹428,624 | ₹3,094 | validation |
| DIXON | 2023-11-13 | 2024-05-21 | 190 | 9.12 | 15.50 | 69.4% | ₹415,396 | ₹3,069 | validation |
| ZYDUSLIFE | 2024-01-23 | 2024-05-21 | 119 | 18.45 | 29.28 | 58.1% | ₹410,755 | ₹3,829 | validation |
| INDHOTEL | 2022-10-17 | 2024-02-26 | 497 | 8.30 | 16.02 | 92.5% | ₹408,599 | ₹2,458 | validation |
| TORNTPHARM | 2023-11-28 | 2024-12-23 | 391 | 79.91 | 131.29 | 63.7% | ₹397,237 | ₹3,443 | validation |
| MCX | 2025-03-03 | 2025-06-09 | 98 | 3.86 | 5.84 | 50.8% | ₹388,629 | ₹4,623 | validation |
| SUNDARMFIN | 2023-05-15 | 2024-08-26 | 469 | 40.53 | 73.07 | 79.5% | ₹387,485 | ₹4,012 | validation |
| Name | Entry | Exit | Days | Entry price* | Exit price* | Return | Net P&L | Costs | Window |
|---|---|---|---|---|---|---|---|---|---|
| HEROMOTOCO | 2024-09-23 | 2025-02-17 | 147 | 13.18 | 8.38 | -36.8% | ₹-348,075 | ₹2,999 | validation |
| BPCL | 2024-09-09 | 2025-02-17 | 161 | 9.05 | 6.45 | -29.0% | ₹-266,022 | ₹3,032 | validation |
| SIEMENS | 2024-12-23 | 2025-04-07 | 105 | 65.07 | 46.75 | -28.5% | ₹-248,916 | ₹2,902 | validation |
| CUMMINSIND | 2024-06-03 | 2025-02-17 | 259 | 41.15 | 30.44 | -26.4% | ₹-215,179 | ₹2,744 | validation |
| INDIACEM | 2024-10-21 | 2025-02-17 | 119 | 7.59 | 5.93 | -22.3% | ₹-206,816 | ₹4,007 | validation |
| CHAMBLFERT | 2025-04-28 | 2025-08-04 | 98 | 28.36 | 21.98 | -22.9% | ₹-201,399 | ₹3,353 | validation |
| COLPAL | 2024-08-26 | 2024-12-23 | 119 | 40.92 | 32.07 | -22.0% | ₹-199,575 | ₹3,124 | validation |
| UTIAMC | 2024-12-23 | 2025-03-03 | 70 | 2.54 | 1.98 | -22.7% | ₹-198,435 | ₹4,791 | validation |
| VTL | 2022-04-11 | 2022-06-13 | 63 | 17.12 | 10.13 | -41.3% | ₹-195,192 | ₹2,262 | validation |
| KIMS | 2025-01-20 | 2025-03-03 | 42 | 3.20 | 2.49 | -22.7% | ₹-192,044 | ₹3,650 | validation |
* Prices are split-adjusted to today by the project's rule, so they can differ from the quoted price on that day. Every trade, fill and cost is in the downloadable files on the review page (trade log and audit tables).
8. The index-direction call
A small LightGBM on market-context features (index returns and moving averages, volatility, VIX level and regime, breadth, drawdown from the 252-day high) predicts whether the Nifty 500 is higher five sessions later. Out of sample, Fridays 2010-2025: accuracy 53.7%, AUC 0.51 (0.50 is chance), against 56.8% for always predicting the majority direction. In validation: accuracy 53.8%, AUC 0.51. A rule that moves to cash on a bearish call reduced validation CAGR from 25.6% to 18.7% for v2. The weekly index call has no demonstrated skill and should not be used.
9. What could make this too good to be true, and the checks
- Small and illiquid stocks. The edge sits where costs are most uncertain. v2 uses a ₹5 crore floor, 0.5% slippage per fill below ₹5 crore and 0.25% up to ₹20 crore; at three times those costs it still earns 19.9% in validation. Real fills in these names should be measured before capital is risked (position size is about ₹1.7 lakh, small against ₹5 crore of daily value).
- Survivorship. With coverage of 74% to 94% in validation and 60% to 71% in development, results are overstated by an unknown amount. With an 8%/yr allowance instead of 4%, v2 validation would still be 17.6% against 11.8% and development 8.0% against 10.1%: validation holds, development does not.
- Size and beta. Portfolio beta to the Nifty 500 is 1.10, and 2020 to 2023 were strong years for small and mid-sized stocks: the year table shows v2 returns of 50%, 61%, 9% and 46% in 2020 to 2023; the gains over the index were smallest in 2018, 2019 and 2022. Part of the "alpha" is a size tilt; random selection from the same universe is the fairer comparison and the model is far ahead of it.
- Execution. Orders are assumed to fill at the next open plus slippage; names that open locked at a circuit limit (zero-range days) are skipped for entry and kept on exit. Partial circuits and auction-session effects are not modelled. Corporate-action adjustment and the 35% jump rule can hide real crashes.
- Overfitting. 540 configurations on development; the reality check, the random null, the plateau table, the shifted sample and one-time validation are the controls. Version 2 and its rule corrections were decided on development data, but only after the v1 development results were seen, and validation was run after the freeze.
- Bug history (disclosed). The first v1 development run misaligned prediction columns with price columns; it was fixed and rerun before any validation result.
10. Recommendation and next steps
- Forward paper run of v2: approved and scheduled. Frozen specification
research/rank500/paper_spec.json; paper start Monday 2 November 2026, ₹50 lakh notional, decisions every second Friday close (first: 6 November), orders recorded for the next open with the v2 cost schedule, a model trained at the start and refitted each January. No orders or broker calls. A nightly job (19:50 IST) replays the run and updates the forward section of the review page. - Sealed period: opened once (section 7b) and now spent for this strategy; it becomes training data for the forward models.
- Measure real execution costs on a small test basket in the ₹5 to ₹20 crore names before sizing up: this needs real orders and therefore your explicit instruction; none have been placed.
- Survivorship gap: work in progress (price histories of delisted and merged members from NSE archives); results are rerun if the data can be recovered.
- Do not use the weekly index call.
11. Sources and files
Prices and membership: Dhan daily bars and NSE Nifty 500 constituent snapshots held in the project (data/research); India VIX and Nifty 500 index: Yahoo Finance series held in data/niftyscalp. Code, pre-registration, results and figures: research/rank500/. Education and research only; not investment advice.
Research and education only. Hypothetical and back-tested results have inherent limitations and do not guarantee future results. Not investment advice; the publisher is not presented as a SEBI-registered investment adviser or research analyst.