All strategies · Longbow Weekly review page

Longbow Weekly: strategy documentation

Version 1.0.1. Frozen 2026-10-19. Owner: Prashant Maheshwari. Status: forward paper run active, no real money, no orders. Everything in this document describes what the code in research/weekly/ actually does. If the code and this document ever disagree, the code is the specification and this document has a bug: report it.

1. One-page summary

2. Data

2.1 Sources

• Prices: daily candles for NSE equities delivered by Dhan through the market-data platform snapshot, merged with extra symbols in data/nse_extra/. Assembled by research/build_panel.py into data/research/panel.parquet (723 symbols, 1999 to the latest session). Quality labels (OK, REVIEW, EXCLUDE) come from quality.csv; the 41 EXCLUDE symbols are dropped.
• Index membership: data/reference/nifty500_constituents.csv: dated snapshots of the Nifty 500 list (effective dates, semi-annual). A stock is a member from a snapshot's effective date until the next snapshot. Because only effective dates are used (never announcements), a newly added stock becomes tradable only after it is actually in the index: slightly conservative.
• Market filter index: Nifty 500 price index. Research used Yahoo Finance's public series; the paper run uses the platform's own series (security id 19). The two agree within 0.04%.

2.2 Cleaning and adjustment (data.py)

• Daily return = close / previous close - 1. Any one-day move bigger than 35% is treated as an unadjusted split, bonus or similar and is removed (set to missing), so it never counts as profit or loss. This also removes genuine crashes of that size (a documented bias in favour of the strategy's results).
• An adjusted price series is built from the cleaned returns (it starts at 1.0 for each stock at the start of the panel, so its absolute level means nothing: only ratios do). Open, high and low are scaled by the same factor. Raw close and raw open are kept for rupee prices, share counts and the price filter in reports.
• Weekly bars: weeks end Friday. Weekly open = first session's open, weekly close = last session's close, high = highest high, low = lowest low, volume = sum.
• Liquidity: average daily traded value over the last 4 weeks = (sum of raw close x volume over the last 4 weeks) / (number of sessions in them) / 10,000,000. Called ADV, in ₹ crore.
• History: number of weeks with a valid close so far.

2.3 Data limits you must know

• Survivorship: only 68% (2016) to 92% (2025) of the stocks that were Nifty 500 members have price history in our records. Missing names are mostly delisted or merged, and many of them did badly. A long-only strategy that picks past winners is flattered by their absence. The back-test applies a 4% a year (stress 8%) haircut on invested capital; this is an estimate, not a measurement.
• Prices are as delivered; the 35% jump rule is a blunt tool.
• The index series is a price index (no dividends). Stock returns are price returns too.

3. The rules, exactly

3.1 Eligibility on a decision date t (end of week t)

A stock can be bought or held if all are true: 1. It is a Nifty 500 member on that date. 2. It has a valid weekly close in week t, and a valid open exists in week t+1 (so it can be traded). 3. ADV over the last 4 weeks is at least ₹5 crore a day. 4. It has at least 60 weeks of price history.

3.2 Score

For each eligible stock, with weekly closes C:

momentum = C[t-4] / C[t-39] - 1          (return from 39 weeks ago to 4 weeks ago; the latest 4 weeks are skipped)
risk     = standard deviation of the weekly returns of the last 26 weeks
score    = momentum / risk

Higher is better. Stocks with a missing score are ignored. Ties are broken by array order (practically never matters).

3.3 Market filter

regime = Nifty 500 weekly close > average of the last 26 weekly closes. * If regime is false at decision week t, no stock is eligible: the target is 100% cash, and every held stock is sold at the next open. This applies in any week, not only on rebalance weeks. * While cash, the account earns the RBI repo rate minus 1.5% a year, spread weekly. (Only used for the idle share of the account.)

3.4 Which weeks are rebalance weeks

Let j0 be the first decision week. Rebalance weeks are j0, j0+4, j0+8, .... For the paper run the first decision is the week ending Friday 2026-10-23, so rebalance weeks end on 2026-10-23, 11-20, 12-18, 2027-01-15, 02-12, 03-12, 04-09, 05-07, 06-04 and so on (a holiday on Friday only changes the date label, not the week).

3.5 What happens on a rebalance week

1. Rank all eligible stocks by score, best first.
2. Keep every currently held stock that is still eligible and ranks within the top 22 (1.5 x 15, rounded down).
3. Fill the remaining slots, from the best-ranked unheld stocks down, until 15 names are chosen or no eligible stocks are left.
4. Give every chosen stock the same target weight of 1/15 = 6.67% of account value, whether or not 15 names were found. Unused slots stay in cash. Winners are trimmed and laggards topped up back to equal weight.
5. Stocks not chosen are sold completely.

3.6 What happens on a non-rebalance week

• Held stocks that are still eligible are kept at their current (drifted) weight: no trimming, no topping up, no new names.
• Held stocks that are no longer eligible (left the index, liquidity fell, price missing, no open) are sold.
• If the market filter is off, everything is sold.

3.7 Execution

• Orders are generated on the decision date's evening and filled at the first session open of the following week.
• Buys fill at open x (1 + 0.2% x cost multiple); sells at open x (1 - 0.2% x cost multiple). The simulation treats the 0.2% as a cost, not as a price change.
• Position sizes are fractional in the simulation. The paper orders_next.csv and positions.csv convert to whole shares at the latest raw close (floor(weight x equity / price)).

3.8 Costs (per side, as a fraction of the traded value)

Buy side about 0.35% (including slippage); sell side about 0.33% plus ₹17.70. All cost items are multiplied by the cost multiple (1x standard, 2x and 3x stress). These are estimates until reconciled with a contract note; Dhan's actual delivery brokerage may be lower.

3.9 What the strategy does not do

No stop-losses (exits are rank-based or filter-based), no volatility scaling, no sector caps, no shorting, no leverage, no options, no intraday trading, no price-impact model, no taxes. Round 2 tested stops and volatility scaling and they added nothing, so they were left out.

3.10 Edge cases

4. How it was built and tested (the history you should know)

4.1 Process

1. Scope. 14 weekly-bar ideas were proposed. Ten could be tested with price and volume data: momentum and 52-week highs, low volatility, trend following (moving average, Donchian, SuperTrend), a price-reaction stand-in for post-earnings drift, base breakouts, market structure, EMA support, anchored VWAP, weekly fair-value gaps, RSI/MACD. Two more were studied as events (index rebalancing) and pairs. Four could not be tested honestly for lack of data: quality and value factors, true post-earnings drift, event-driven and merger arbitrage, volatility dispersion.
2. Windows. Development 2016-01-01 to 2025-09-30, split into train (2016 to 2021) and validation (2022 to 2025-09). A hold-out from 2025-10-01 was sealed.
3. Round 1. 597 combinations (idea variants x three market-filter settings), each at 1x/2x/3x costs and 0/4/8% survivorship haircut. Screen: CAGR above 10% and worst fall under 20% at 2x costs plus 4% haircut. One combination passed.
4. Round 2. 288 focused momentum combinations: four ranking types, 10/15/20 names, four market-filter versions, stock trailing stop (none, 20%, 30%), volatility scaling (off/on). 25 passed both periods. They cluster on one plateau: risk-adjusted momentum (39 or 52 weeks, skip 4), 10 to 20 names, filter on the 26- or 30-week average. Stops and volatility scaling did not help, so the simplest version was frozen: risk-adjusted 39-week, 15 names, filter on the 26-week average.
5. Checks on the chosen rule. Cost and haircut stress, start-date sensitivity, two random-pick nulls, a walk-forward re-selection test, and White's reality check.
6. Hold-out. Opened once on 2026-10-19 at the owner's explicit instruction (ledger: holdout_ledger.json).
7. Paper run started the same day.

4.2 Development results (candidate, 2x costs, 4% haircut)

4.3 Stress (full development window)

4.4 Calendar years (2x, 4%)

Two years (2021, 2023) do most of the work. This is the single most important weakness to keep in mind.

4.5 Other checks

• Start-date sensitivity: starting in 2016, 2017, 2018, 2019 or mid-2020 gives 21% to 36% a year but the worst fall is 17% to 25% (22% from 2017 on). "Under 20%" is not robust.
• Null 1 (random 15 stocks, same filter, 60 runs): -0.1% a year on average. Null 2 (random 15, no filter): -8.9%. The ranking has information; the filter alone does not make money.
• Walk-forward (re-pick the best round-2 setting each year from earlier data, 2019 to 2025-09): 22.8% a year with a 23.3% worst fall.
• Reality check (White) across all 288 round-2 variants on the validation period: best t-statistic 2.85, p = 0.047. Borderline; roughly 900 variants were examined in total in the study, so some selection bias remains.
• Rolling 54-week windows in development (456 windows): negative in 9% (standard costs) or 19% (2x costs + haircut); worst window -10.2% or -14.7%; median +22.1% or +15.2%. When the index lost money over 54 weeks (82 windows), the strategy's median result was +3.7% at standard costs and -1.6% at 2x costs.

4.6 The hold-out (opened once)

Market filter was on in 56% of weeks (cash in 44%). The candidate beat 98% of the random runs. Monthly returns at standard costs: Oct 2025 +2.2, Nov +1.1, Dec -1.2, Jan 2026 -6.4, Feb +0.3, Mar +0.3 (index -9.3), Apr +0.4 (index +7.9), May -1.6, Jun +0.3, Jul -4.6, Aug +7.9, Sep +0.4, Oct +0.2.

Correction added 2026-10-20: the hold-out total depends on the rebalance calendar. The strategy rebalances every fourth week, so the result depends on which of four weekly "phases" the cycle starts on. The -1.4% above is for the phase set by a 2016-01 start. Re-running the same rule with the cycle shifted by one, two or three weeks gave +5.3%, +0.6%, +1.5% and -1.4% over the same 55 weeks (standard costs): a range of about 7 points, all better than the index's -3.3%. So "did not confirm a profit" overstated the case: the honest statement is "between -1.4% and +5.3% at standard costs depending on phase, still far from a clear confirmation". Phase luck is a real property of any monthly strategy. A possible future design (a new version, not applied here) is to split the account into four equal sleeves that rebalance in different weeks, which removes the phase choice. The forward paper run uses one fixed phase (first rebalance 2026-10-23).

Reading: the ranking showed skill against random picks and the filter limited the fall, but there was no profit, and 55 weeks cannot separate skill from luck. The hold-out is now spent: it cannot be re-run with changed settings.

4.7 Other ideas, in one line each

Low volatility, trend following, anchored VWAP, EMA support, market structure, fair-value gaps and RSI/MACD had raw returns of 5% to 30% a year but drawdowns of 33% to 53% (or little return) at 2x costs: none passed the drawdown screen. Base breakouts lost money. Index-rebalancing effects were worth roughly 2% to 3% a year at best. Pairs showed no edge after costs.

5. How the paper run works

• Script: research/weekly/paper.py, started by research/weekly/nightly_paper.sh (cron, Monday to Friday 19:15 IST), after the 18:45 job has refreshed data/research/panel.parquet. Then the review site is rebuilt.
• Deterministic replay. Each run re-simulates the whole paper account from the start date with the same engine as the back-test. State cannot drift, and a missed night is repaired by the next run.
• Completed weeks only. The last weekly bar is used only once its week is over (Friday session present, or today is past the week's Friday).
• Placeholder week. To obtain the decision at the last completed week (whose trades happen at the next open), the script appends a placeholder week with the last close as price, runs the engine, and drops that week from the equity. (A first version used empty prices and wrongly produced "sell everything"; fixed 2026-10-19 before any live decision.)
• First decision: week ending 2026-10-23 (after the 19:15 run that evening). First simulated fills: 2026-10-26 open.
• Outputs in data/paper/weekly_mom/ and on the review page:

• Frozen specification: research/weekly/paper_spec.json. No parameter may change during the run.
• Known simplifications: fills at the open plus a fixed 0.2% (no partial fills, no circuit limits, no price impact); fractional shares in the equity curve; costs at 1x.

6. How to read the results for a decision

6.1 Principles

1. The back-test is a hypothesis. The paper run is the only clean out-of-sample evidence left.
2. Judge on weeks and rebalances, not days. Nothing before 26 weeks and 7 rebalances is evidence.
3. A good back-test followed by a weak hold-out is the typical signature of optimism in the back-test. Do not respond by tuning on the hold-out; that turns it into training data.
4. One year of a 15-stock portfolio has ±15% of pure noise. A single result, good or bad, proves little.

6.2 Check dates

• Early look: after the 2nd rebalance (week ending 2026-11-20): is everything working as described? Look for operational problems only; do not judge performance.
• 26-week review: decision week ending about 2027-04-23 (7 rebalances).
• 52-week review: decision week ending about 2027-10-22 (14 rebalances).

6.3 Pass, hold and stop criteria (written now, before any paper result)

Measured on the paper equity curve at standard (1x) costs, versus the Nifty 500 price index and cash (repo minus 1.5%, about 4.5% a year):

Passing both reviews means permission to consider a small live pilot, not an instruction to go live.

6.4 If it passes: a cautious live path (owner decides)

• Start with about a quarter of the intended capital; same rules, no tuning.
• Place orders manually in the first 15 minutes after Monday's open using limit orders within about 0.5% of the open; log every fill and compare with the paper price.
• Run live and paper in parallel for at least 13 weeks; slippage in live trading above about 0.4% a side would mean the cost assumption is too low.
• Add capital only in steps, after each further 26 weeks that meet the pass bar.

6.5 What the numbers cannot tell you

• How much survivorship flattered the history (the 4% and 8% haircuts are guesses).
• Real liquidity and spreads for small-cap names at your order size.
• The effect of taxes: holdings are usually kept 1 to 6 months, so gains are mostly short-term capital gains (taxed at 20% at the time of writing; confirm current rates and your own position with a tax professional). The back-test has no tax.
• Whether momentum will keep working: it is a documented market effect but it has bad years and reversals (for example January 2026 in the hold-out).

7. Weekly operating routine (for the paper run today and a live pilot later)

Friday or Saturday morning 1. Open the strategy-review page or events.log: check that the latest line shows the data date of this Friday (or the last session) and no "stale" warning. 2. Read orders_next.csv. If empty: nothing to do. If it has SELL or BUY lines: these are the next open's orders. 3. Read signals_latest.csv for context (top 15 are the targets).

Monday, after 9:15 open 4. Paper: nothing to do, the replay fills at the open price automatically next evening. 5. Live pilot (later): place the sells first, then the buys, using limit orders near the open; record price and quantity in a log.

Monday evening 6. Check that positions.csv matches the broker positions (live pilot) and that cash is as expected.

Monthly 7. Compare equity with the index; note months where the filter was off.

Never 8. Change a rule because of one bad or good week. Add a stock by hand. Skip a SELL because "it will recover".

8. Change control

• The specification is frozen for the duration of the run (paper_spec.json, version 1.0.1).
• Any change to a rule, parameter, universe, cost assumption or filter means a new version number, a new frozen spec, and a restart of the paper evidence (the old paper results stay on record).
• Bugs that change behaviour away from this document are fixed and recorded in section 10; they do not count as strategy changes if they restore the documented behaviour.
• The hold-out for this strategy is spent. Further out-of-sample evidence comes only from the paper run (or live).

9. Risks, honestly

10. Known issues and fix log

• 2026-10-19: SuperTrend variant in the research families produced no signals (NaN trap); fixed and the grid re-run. Did not affect the chosen strategy.
• 2026-10-19: adjusted prices start at 1.0 per stock, which broke a rupee price filter in a separate daily study; the weekly strategy only uses ratios, so it was unaffected.
• 2026-10-19: the paper run's placeholder week had no prices, so the final decision liquidated everything in orders_next.csv. Fixed (placeholder repeats the last close) and re-tested before the first live decision: the one-week case now gives 15 BUY orders; a later non-rebalance week gives none.

11. Questions people ask

Why weekly and not daily? Weekly bars cut noise, cut turnover, and match how momentum is usually measured. Daily versions were not tested for this strategy.

Why skip the latest 4 weeks? Stocks that have just risen sharply tend to give back part of the move over the next month (short-term reversal). Skipping that month is the standard "12-1" convention (here 9-1).

Why divide by volatility? Two stocks with the same return are not equally good if one got there smoothly and the other lurched. Dividing favours the smooth one. It was better than raw momentum in testing (Round 2).

Why the 26-week market filter? It keeps the account out of broad sell-offs (for example 2020). It costs missed rebounds.

Why 15 stocks? A compromise: enough to diversify, few enough for the ranking to matter; 10 and 20 names gave broadly similar results (10 had larger falls, 20 slightly lower returns).

Why monthly rebalancing? Weekly rebalancing raised turnover by about half (6.0 versus 4.0 times a year in the grid) and did worse after costs.

Will it make money next year? Unknown. The development result was strong; the hold-out year was slightly negative at standard costs and positive relative to the index. Only the paper run can add evidence.

Is this advice? No. It is research and education. Not investment advice; the publisher is not a registered investment adviser or research analyst.

12. Where everything is

Reproduce the development numbers: python research/weekly/run.py, round2.py, final.py (development data only). Do not run holdout.py again: it refuses to.

13. Execution audit: how to confirm trades follow the signals (added 2026-10-20)

What you can check, and where. The review page has an "Execution audit" section and a trade log with these columns: signal date (the Friday close when the ranking was computed), entry fill (the next week's first session), entry price, rank on the signal date, exit signal, exit fill, exit price, and the exit reason. The downloadable fills_*.csv lists every transaction (including top-ups and trims) with signal rank and score.

Checks run automatically on every build (they can be re-run from the CSV files): 1. Every transaction happens after its signal (fill date later than signal date; the typical gap is 3 days, Friday to Monday). 2. Every new purchase is for a stock ranked inside the target list on its signal date (new entries rank 1 to 15; top-ups of names already held can rank lower). 3. Every full exit has a documented reason: rank fell below the keep limit (22), no longer eligible, or the market filter turned off. 4. Every execution price is a real traded price: the stock's open on the fill date lies inside that day's low to high. 5. Never more than 15 names held at once, and entry dates precede exit dates. 6. The trade ledger reconciles to the equity curve: an independent rupee accounting of every fill (cash, positions, costs paid from cash) ends within about 1.4% of the simulator's equity.

Extra proof for the forward paper run. * One signal snapshot file per decision week (paper_signals/<name>/signals_log/YYYY-MM-DD.csv) is written the evening of the decision and never overwritten. It holds the top-40 ranking with scores, the target weights and the action for each name. * signals_ledger.jsonl chains each file's SHA-256 hash to the previous one with the time it was generated. The review page has a button that re-hashes every file in your browser and compares the chain. * The nightly job commits the snapshots to GitHub; GitHub's server time on the commit is an independent timestamp that the signal existed before Monday's open. * Every night the system re-derives each past decision from the data and compares it with the stored snapshot. Any difference is reported as a failed check. * Check "each snapshot was written before the session in which it was traded opened" compares the ledger's generation time with 09:15 IST on the fill date. * Limits. A paper run proves internal consistency and timing; it cannot prove that real orders would have filled at those prices. When trading real money, reconcile these logs with the broker's contract notes. Tests that replay history (not generated live) correctly show the timing check as failed: only live-generated snapshots can pass it.

Fix log. 2026-10-20: the first ledger version used the signal date's own session as a possible fill date for holiday-shortened weeks; fixed so only sessions strictly after the signal count (the audit caught this: it failed the "after its signal" check until fixed).

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.