Strategies
Longbow Learned v1.0.0
Machine-learning stock ranking blended with the Longbow rule (TensorTrade-inspired study) · A LightGBM model that estimates which stocks will land in the top fifth of the next 4 weeks' returns, averaged 50/50 with the Longbow momentum score; same portfolio rules as Longbow Weekly.
Promising but unproven: hold-out showed no difference. The one-time check on 2025-10 to 2026-10 (run 2026-10-20 under pre-registered rules) found no reliable difference between the machine-learning blend and the plain rule: +5.4% against +5.3% at standard costs (+0.8% against +0.5% at 2x costs plus the 4% allowance), with the index at -3.3%. The development advantage (30.8% against 22.1% a year) is intact but not reinforced. The forward paper run continues.
Forward paper run: 1 sessions, equity ₹2,000,000
Longbow Weekly v1.0.1
Fourteen weekly Nifty 500 ideas tested; risk-adjusted momentum with a market filter is the only candidate · Weekly scan of liquid Nifty 500 stocks: top 15 by risk-adjusted 39-week momentum (skipping the last 4 weeks), monthly rebalance, in cash when the Nifty 500 is below its 26-week average.
Candidate: hold-out did not confirm profit. The sealed hold-out (2025-10 to 2026-10-16, opened once on 2026-10-19) did not confirm a profit: -1.4% at standard costs against -3.3% for the Nifty 500, with a 13.8% worst fall. It protected capital in a down year but did not make money. The forward paper run continues.
Forward paper run: 1 sessions, equity ₹500,000
Meridian v1.0.0
Nifty 500 Momentum · Nifty 500 cross-sectional momentum, regime-gated
Hold-out did not confirm. Strong in the development window, but the 12-month sealed hold-out lost money and trailed the index; the forward paper run is now the deciding evidence.
Nightjar v1.1.0
Overnight short strangle on Nifty weekly options, re-costed with standard Dhan charges and 1% slippage per full cycle · The same pre-registered finalist as Nightjar 1.0, re-tested after the owner replaced the cost assumptions with the standard Dhan charge schedule and 0.5% slippage per fill (1% per full cycle).
Passes the re-costed rule; forward paper run scheduled from 2026-10-26. Same finalist, cheaper standard costs: Nifty +₹1,476 per night on the sealed period (t 2.6), +₹1,681 overall; Sensex +₹693 (t 1.8) but only +₹344 in its second half.
Forward paper run: waiting for the first session
Meridian Adaptive v1.1.0
Nifty 500 Momentum, risk tiers · Same engine with breadth- and volatility-based risk tiers
Rejected. Higher headline return, but better than the standard version in only 4 of 10 years with no improvement in drawdown.
Cinder v1.0.0
ATM EMA short seller on frozen Nifty weekly options (EMA2/EMA20 cross on the 5-minute premium chart) · Sells the frozen ATM call or put when its 5-minute EMA2 crosses below EMA20, with a 20% stop, an 80% premium-drop target, an EMA10 trail and a 15:15 square-off.
No edge: loses before and after costs. Development 2277 trades: mean ₹-699 per day (t -6.3), win rate 24%, gross ₹-81 a trade before ₹130 of charges; validation ₹-593 a day. Random-time shorts with the same exits: median ₹-640 a day.
Compass v0.2.0
ML direction model for Nifty options (brief ML-NIFTY-F&O-01), Gate 1: does ML beat buy and hold? · Tests whether machine-learning models predict Nifty's 1 to 5 day direction better than simply staying long, out of sample over 2013 to 2024, before any options work.
Gate 1 failed: no skill over buy and hold. Across 12 pre-registered variants the best machine-learning model gained +0.006% per decision over always-long (t = 0.06, reality-check p = 0.87). The project stops here by its own rule.
Compass A v0.2.0
Do option-market features (IV, skew, put/call open interest) add direction information for Nifty? · Tests whether implied volatility, skew, put/call open interest and implied move predict Nifty's next-session direction beyond price and VIX features.
No detectable skill at this sample size. Best option-feature variant gained +0.057% per session over the better baseline (t = 0.58, p = 0.26); reality-check p = 0.37 across 8 variants.
Doji Ember v0.1.0
Nifty 500 daily doji + 5/20 EMA, OTM options · Doji bar in an EMA5/EMA20 trend with a 1-to-2 directional streak: buy one-strike OTM call (or put), exit on EMA5 close, 2x, or day 5.
No robust edge. The result flips from +3.4% to -5.9% per trade depending on one fill assumption, so there is no reliable edge. Calls never make money; puts only in the optimistic case.
MGTS-v1D v1.0.0-D
Gamma-aligned multi-timeframe option spreads on Nifty weekly options (daily open-interest gamma, not live GEX) · Classifies the day by dealer gamma (rebuilt from the previous close's open interest), then sells spreads at prior-day and opening-range extremes in positive gamma or buys breakout spreads in negative gamma.
No edge, and the idea cannot be tested as written. Development 137 trades, mean ₹-17 per day (t -1.4); validation 28 trades, ₹-25; 3 of 11 conditions passed
Nightjar v1.0.0
Overnight short strangle on Nifty weekly options, inspired by a Tradetron Sensex report · Tests whether selling a short strangle at the end of the day and buying it back the next morning keeps its profit once realistic costs are charged.
Passes on Nifty, fails the Sensex replication: not paper-traded. A pre-registered finalist (sell within two sessions of expiry at 15:15, buy back next midday) made +₹871 a night at 2x costs on 50 sealed Nifty nights and +₹1,035 over all 196 nights, but on Sensex it was +₹368 then -₹331 (about zero overall). Rule set beforehand: both halves of Sensex positive. It failed, so no paper run.
Obsidian Atlas v1.0.0
CRUDEOILM end-of-day search · 264 configurations with reality-check p-values and sub-period stability.
No candidates. None of 264 configurations passed the reality-check and stability gates.
Obsidian Nocturne v1.0.0
CRUDEOILM sell-today-buy-tomorrow · Short at the close, cover next morning (STBT), with the mirror long (BTST).
No edge found. The market shows no overnight drift to harvest, so the ₹58 round-trip cost makes STBT lose money on average, and the ₹600 maximum-loss rule cannot be enforced overnight.
Obsidian Sentinel v1.0.0
CRUDEOILM daily signal · Daily-bar signal with intraday execution, walk-forward.
No edge found. Out-of-sample t-statistic -0.83 over 80 trades; nothing to promote.
Obsidian Vector v2.0.0
CRUDEOILM reversion, gap, shock · Reversion, gap, volatility-shock and weekday variants.
No edge found. Out-of-sample t-statistic 0.00. Volatility-shock momentum is the only recurring lead and stays a hypothesis.
Quickfuse v1.1.0
Nifty 50 weekly-option fused scalp (VWAP, ORB, EMA, CPR, RSI, price action, volume) · Buys an ATM weekly call or put when 4 of 5 trend, pullback, momentum, volume and breakout conditions line up, in two intraday windows.
Closed: no edge. After 186,480 tested combinations, no setup is profitable and robust at 2x costs with drawdown under 20%. The specification as written loses ₹268 a trade (t -4.9), no better than random entries (-₹240).
Quiet Harbour v1.0.0
Buying statistical dips inside a regression uptrend (owner-supplied code) · Buy when a 100-bar regression slope is positive with R-squared of at least 0.35, the 20-bar z-score is at or below -1.5, and price has not crashed more than 2.5 ATR; stop 2 ATR, target 3 ATR, risk 0.5% of equity per trade.
No edge. As supplied the rule earns almost nothing: 0.2% a year on the Nifty 50 index and about 0% a year on a Nifty 500 stock portfolio before stress. At 2x costs the stock portfolio loses 7% a year. No variant of 144 clears 8% a year with drawdown under 20% in both periods.
Tidewatch v1.0.0
Nifty multi-day futures and options (trend, momentum, breakout, reversion, VIX, calendar) · Tests whether holding Nifty for days or weeks, where costs are small next to the moves, gives a profitable strategy with drawdown under 20% of ₹5 lakh.
No robust edge; passes the screens only at ₹15 lakh. At ₹5 lakh nothing passed. At ₹15 lakh, where one futures lot is about 1x leverage and the drawdown limit is ₹3 lakh, 17 of the 44 rules are profitable at 3x costs in both periods within that limit, but the gains are small (about 1% to 5% of capital a year above cash), the trend rules show no timing skill, and the reality check across them gives p = 0.175.
Tidewater v1.0.0
RSI(14) long reversal after a 50-day stretch below 50, Nifty 500 stocks, ₹10 lakh, 25% drawdown limit · Buys when a stock's RSI crosses above 50 after more than 50 calendar days below it, with a stop under the 50-day low and three profit exits.
No edge: failed the pre-registered rule. Net CAGR +5.0% in development (647 trades, win rate 56%, average +0.10R), -3.3% in validation; alpha versus the Nifty 500 +1.4% a year (t 0.5) and negative at 2x costs. Random entries with identical exits earned a median +5.5%.
Tidewater R v2.0.0
Regime-scaled risk: 2% at risk in a calm uptrend, 0.75% otherwise (separate return-boost test) · Tests whether risking more when the market is above its 200-day average and the VIX is calm, and less otherwise, lifts the RSI-50 reversal strategy's returns.
Failed: more risk, same lack of edge. Regime-scaled risk gave +0.8% in development (drawdown 30.5%, breaching the 25% limit) and -2.6% in validation, below the random-entry medians (+4.0% and -4.9%).
Tidewater v2 grid v2.0.0
27 variants: RSI level and days below it (30/40/50) with no, trend (200-day) or volatility (VIX) regime filters · Tests whether other RSI levels and day counts, and a market-regime filter, rescue the RSI-50 reversal rule; one pre-registered run on development, selection by development only.
No variant or filter helps. Selection on development alone picked the original 50/50 rule without a filter: net +5.0% in development, -3.3% in validation; reality-check p 0.89 across the 27 variants (the best raw t-statistic belongs to a variant with one trade).
Compass B v0.3.0
Short-horizon cross-sectional ML ranking, long/short in stock futures (development gate) · Tests whether ranking stocks by a LightGBM score at 1 and 3 days and going long the top and short the bottom survives futures-level costs.
Failed the lot-aware gate at the stated costs; unproven at realistic costs. At 1% slippage per full cycle the 5+5 book on Rs 20 lakh loses -318% of capital over 2024 to Sep 2025 (t = -5.2). Break-even slippage is 0.28% per full cycle; liquid stock futures typically cost far less, but that is unmeasured.
Obsidian Pincer v0.1.0
CRUDEOILM inverted strangle · Both option legs in the money; needs options data we do not yet hold.
Not testable yet. The bhavcopy files we hold contain futures only, so no option price, implied volatility or margin can be tested; we will not invent premiums.
Ironbark v1.2.0
Nifty 500 five-day LightGBM swing with TabFM veto and ATR sizing · Tests a LightGBM five-day ranker (with a Google TabFM veto idea), ATR sizing and a drawdown halt, first as delivery equity, then as stock futures in whole lots at ₹50 lakh.
Signal real, edge did not hold on the second look. Delivery equity fails on costs. As stock futures in whole lots at ₹50 lakh, the 4-name version made +13.0% a year at 2x costs in development (2018-2023, drawdown 7%) but only +4.7% in 2024-01 to 2025-09 (index +9.4%, drawdown 19%); the 6-name version made -1.4%. Pre-registered gate failed; paper run cancelled; hold-out left sealed.
Research log
| Date | Strategy | Event |
|---|---|---|
| 2026-10-21 | Nightjar | New: Nightjar, overnight short strangle study inspired by a Tradetron Sensex report. Gross edge exists but is thin after costs; a pre-expiry finalist passed the one-time Nifty check and failed the Sensex replication. No paper run. |
| 2026-10-21 | Ironbark | Closed: lot-aware futures version at ₹50 lakh failed the pre-registered second look (4 names +4.7% at 2x vs index +9.4%); paper run cancelled; full equity, trades, fills and audit checks published. |
| 2026-10-20 | Ironbark | Follow-ups: intraday version fails; futures-cost version passed the one-time test (4 names 2x +9.6%, drawdown 6.4%); notional forward paper run starts 2026-10-21. |
| 2026-10-20 | Ironbark | New: Ironbark (MFS-15, LightGBM + TabFM swing, ₹20 lakh) tested on 2018-2023: signal real (IC +0.05 to +0.065), gross +29% a year, net negative at standard and 2x costs. Not paper-traded. |
| 2026-10-20 | Quickfuse | Re-tested at ₹15 lakh (buying and selling grids, walk-forward): conclusion unchanged; walk-forward ORB +₹79,800 at 1x, -₹57,400 at 2x, -₹194,600 at 3x. |
| 2026-10-20 | Tidewatch | Re-tested at ₹15 lakh: 17 of 44 rules profitable at 3x costs in both periods within a 20% drawdown, but excess returns are small (about 1% to 5% a year) and not significant (reality check p 0.175). |
| 2026-10-20 | Quickfuse | Short side tested at the owner's request: the strategy already traded puts on short signals (calls -₹262 a trade, puts -₹274). Option SELLING (35,200 single-leg and 27 straddle combinations, 1x/2x/3x costs) also fails: 0 survive 2x with drawdown under 20%. |
| 2026-10-20 | Longbow Weekly | Correction: Longbow Weekly hold-out total depends on rebalance phase (-1.4% to +5.3% at standard costs); index -3.3%. Trade logs, fills and execution audit added to the back-test and validation segments. |
| 2026-10-20 | Longbow Learned | One-time check on 2025-10 to 2026-10 under pre-registered rules: ML blend +5.4% vs rule +5.3% (standard costs), +0.8% vs +0.5% (2x + 4%): no reliable difference. Rebalance-phase sensitivity found: totals range -2.6% to +5.4% (ML) and -1.4% to +5.3% (rule) by cycle start. |
| 2026-10-20 | Longbow Learned | TensorTrade-inspired study on Nifty 500 weekly: ML ranker blended with the Longbow rule 30.8% a year / 19.8% worst fall out of sample vs 22.1% / 22.0% for the rule (reality-check p 0.072); RL exposure agent failed. Forward paper run started (₹20 lakh, no orders). |
| 2026-10-19 | Longbow Weekly | Hold-out opened once at the owner's instruction (2025-10 to 2026-10-16): -1.4% at standard costs vs Nifty 500 -3.3%, worst fall 13.8%; negative under stress. Did not confirm profit. Paper run continues. |
| 2026-10-19 | Longbow Weekly | Forward paper run of the weekly momentum candidate started (frozen spec, ₹5 lakh, no orders). First decision 2026-10-23 close, first fills 2026-10-26 open. |
| 2026-10-19 | Longbow Weekly | Fourteen weekly Nifty 500 ideas studied (10 testable, 4 parked for missing data). Risk-adjusted momentum with a market filter is a provisional candidate: 24.8% a year, 16.7% worst drop at 2x costs, p 0.047. Hold-out sealed. |
| 2026-10-19 | Quiet Harbour | Owner-supplied mean-reversion-in-uptrend code reproduced and tested: no edge (stock portfolio 0% a year at 1x costs, -7% at 2x; 144-combination grid, none clears 8% with drawdown under 20%). |
| 2026-10-19 | Tidewatch | Multi-day Nifty futures and options tested: 44 rules, 2008-2022, costs 1x/2x/3x. No rule is profitable at 3x with drawdown under 20% of ₹5 lakh in both periods; turn of month is the only consistent one but drew down 32% in 2017-2022. |
| 2026-10-19 | Quickfuse | Closed. Forward paper run cancelled before its first session at the owner's instruction. A follow-up scan of 108 Nifty-futures mechanisms found no edge either (reality check p 0.70). |
| 2026-10-19 | Quickfuse | Multi-combination search (255,600 combinations at 1x/2x/3x costs, capital ₹5 lakh): no strategy is robustly profitable at 2x with drawdown under 20%. ORB family is the only lead; walk-forward +₹31k at 1x, -₹19k at 2x. Hold-out still sealed. |
| 2026-10-19 | Quickfuse | Fused Nifty weekly-option scalp tested on 328 sessions: 863 trades, -₹260 a trade after costs, no better than random entries. Hold-out sealed. Forward paper run started 2026-10-19 at the owner's request. |
| 2026-10-18 | Doji Ember | Daily doji + EMA5/20 option strategy tested on Nifty 500 (17,900 trades, 2016-2025): +3.4% per trade with optimistic fills, -5.9% with conservative fills. No robust edge; hold-out sealed. |
| 2026-10-18 | Obsidian Nocturne | STBT / BTST overnight tests on 748 nights: no edge (best family-wise p 0.993), ₹58 cost makes STBT lose ₹63 a night, gaps over ₹600 on 8% of nights. |
| 2026-10-18 | Obsidian Atlas | Re-run of the 264-configuration search on MCX bhavcopy data: still 0 candidates. |
| 2026-10-18 | Obsidian Pincer | Inverted strangle parked: option price history is needed (bhavcopy supplied contains futures only). |
| 2026-10-18 | Meridian | Sealed hold-out opened once (2025-10-01 to 2026-10-16): -5.2% vs index -2.9%, win rate 30%, gross P&L negative. Did not confirm the back-test; forward paper run continues. |
| 2026-10-18 | Meridian | Slippage stress added (0.5%, 1%, 2% per side) and broker cost models (Zerodha, Dhan, Kotak Neo, Anand Rathi). |
| 2026-10-18 | Meridian | Specification frozen (v4 risk model, 182/7 days, 15 names, 4% each). Paper run scheduled from 2026-10-19. |
| 2026-10-18 | Meridian Adaptive | Rejected: better than v4 in 4 of 10 years and no drawdown improvement. |
| 2026-10-18 | Meridian | Round-2 robustness: sub-periods, concentration, parameter sensitivity, random-rank null (30 seeds). |
| 2026-10-18 | Meridian | First back-test on the point-in-time Nifty 500 panel (723 symbols), ₹5,00,000, 2016-01 to 2025-09; hold-out kept sealed. |
| 2026-10-18 | Obsidian Atlas | MCX bhavcopy history loaded (52 contracts, 2023-04 to 2026-10) for a longer end-of-day re-test. |
| 2026-10-16 | Obsidian Atlas | 264 configurations tested with reality-check and stability gates: 0 candidates. |
| 2026-10-16 | Obsidian Vector | Reversion, gap, volatility-shock and weekday variants: out-of-sample t = 0.00. |
| 2026-10-16 | Obsidian Sentinel | Walk-forward 2023-03 to 2026-01: t = -0.83 over 80 trades. |