All strategies · Review page

Monthly Bollinger Band breakout on Indian equities at ₹50 lakh: build, search and test, 2006 to September 2025

Specification fixed 1 Nov 2026 before any result (research/bbm/STRATEGY_BBM.md, with dated amendments) · NSE data for every name that traded, delisted ones included · no leverage, whole shares, Indian delivery charges and liquidity-tiered slippage · research only; the period from October 2025 was never loaded · 211 configurations tried

1. What was built

• Signal (month t close): close above SMA(20) + 2 × standard deviation (population) of the last 20 monthly closes, and no such close in the six months before (the raw band test, so filters do not move the "first breakout" definition). Entry at the open of month t+1.
• Baseline exits (evaluated at month-end, sell at the next open): close below the 20-month SMA, or Supertrend(10, 2) flips from above to below (ATR Wilder 10 on monthly bars, standard band carry-forward).
• Entry filters tested: RSI(14, Wilder, monthly) at or above 60, 65 or 70 on the qualifying month's close; real body (|close − open| ÷ (high − low)) at least 50%, 60% or 70%, bullish only and any direction; range (high − low) ÷ ATR(14, monthly, through month t) at least 1.0, 1.618, 2.0, 2.618, 3.0; and their pairs and triples.
• ATR exits tested: stop = previous month's close − k × previous month's ATR(14), k = 1.0 to 3.0, in two distinct forms: intramonth (on daily bars, fills at the stop, or at the open when the session opens below it; gap losses recorded) and month-end (monthly close below the level, sell at the next open); each as an addition to the baseline exits and as a replacement for them.
• Portfolio-wide exit: at a month-end, if more than T Nifty-200 stocks have a qualifying signal and at least ceil(75% × open positions) show an unrealised gain of 8% or more (price only, against average cost), sell everything at the next open; no new entries until the earlier of two calendar months after the exit date or the Nifty 200 closing 10% below its all-time closing high (running maximum since 2004); signals produced during the block are dropped.
• Capital: ₹50 lakh, no leverage, whole shares at the unadjusted open price, cash earning repo minus 1.5%; competing signals ranked by liquidity; sizing, position and sector limits, drawdown control and regime filter tested one at a time and in combination.

2. Data, universe and bias controls

• Survivorship: prices from the NSE daily file for every EQ/BE/BZ name 2004-2026 including delisted and merged names; the universe is point-in-time (the 500 most traded names each month, ₹5 crore median daily value at the signal date), a stand-in for the Nifty 500 list (57% to 81% overlap with the real lists). A held name whose prices stop for 3 months is closed at its last price less a 30% haircut (2 cases in the finalist run).
• Corporate actions: the previous-close field in the NSE file is usually not adjusted on the ex-date (found during this build: Reliance's 2024 bonus appeared as a −50% day). Unadjusted split and bonus days are now detected (1,491 days) and removed. This is a heuristic; rights issues and demergers are not adjusted. Dividends are excluded for the strategy and for the index (price index) alike, which understates both by a similar amount.
• Look-ahead: signals use only data through the month-end close; orders at the next first-session open; ATR and RSI include the signal month, available at its close. Index context (the regime filter, the all-time high) uses data through the same close.
• Sectors: NSE's industry column for the current Nifty 500 (20 groups); delisted names are unclassified and are not sector-capped, so the sector cap understates concentration before 2015.
• Windows: development = signals to December 2016; validation = 2017-01 to 2025-09; the period from October 2025 was never loaded. Any statement about the finalist in 2006-2016 is in-sample.

3. Results for the baseline and the finalist

Baseline (as specified, 20 names, no filters), cost stress and windows

Finalist: rsi70+body60+rng1.0 | ATR month-end + base k1.0 + FILL N20 cap10 + regime sma10 entry + RISKFILL 1% N20 cap10

Selected on development data only, at 2x costs, by the pre-registered rule: highest Sharpe among 17 configurations passing the gates (at least 60 trades, drawdown at most 35%, CAGR above the index, at least 60% of capital deployed on average). It is the best of 211 configurations, so its development figure is optimistic by construction.

Full-period detail (finalist, 1x costs, ₹50 lakh)

Baseline for comparison (1x costs)

Costs and gap risk (1x)

The finalist uses month-end stops, so it has no intramonth gap loss; the intramonth ATR variants in section 5 show gap losses of ₹6 lakh to ₹25 lakh over the whole path. Slippage tiers: 0.05% a fill at ₹100 crore a day traded or more, 0.10% from ₹20 crore, 0.25% below (spread included), × 1, 2 and 3 in the stress rows.

4. Is the edge real? Overfitting and multiple testing

Random entries (same number of positions each month, drawn from the same eligible stocks, same exits and sizing) earned a median of 5.1% in development and 9.3% in validation (90th percentiles 9.3% and 14.8%). The finalist's development result is above the 90th percentile; its validation result is above the 90th percentile of 14.8%. The sign-flip reality check asks whether the best of 211 configurations is better than the best you would get from noise: it is not (p = 0.70). Walk-forward re-selection each January gave 13.6% a year against 14.0% for the index.

Plateau and sensitivity

78% of the neighbours beat the index in both periods. Start-date sensitivity (fresh start from cash, 2x costs, full path to 2025-09): 2006: 16.5% against 11.3%, 2008: 12.5% against 8.2%, 2010: 15.6% against 10.7%, 2012: 17.0% against 13.9%, 2014: 17.9% against 13.3%, 2016: 15.6% against 12.9%. Other sensitivities: look-back counts only filtered breakouts: 12.0%, drawdown 34.9%, ATR through t-1 for the range filter: 16.5%, drawdown 30.8%, random signal order (mean of 10): 16.3%, drawdown 32.9%, slippage x1: 16.6%, drawdown 34.7%, slippage x3: 15.7%, drawdown 35.4%.

Parameter combinations tested

Every configuration, with returns, drawdown, win rate, profit factor, expectancy, holding period, capital deployed, positions, open risk, recovery time, worst month and year, losing streak, costs, gap loss and missed signals, is in configs.csv on the strategy page. The ten best by development Sharpe among the combinations (selection data) and how they did afterwards:

5. What each choice does

Entry filters

Baseline 8.0% in development and 9.9% in validation (2x costs). 79% of the 81 entry sets beat the baseline in validation but only 15% did so in development: the filters mostly helped in 2017-2025 and not in 2006-2016, so no filter is robustly better. Strong-candle filters and wide-range filters cut trade counts and capital use (a range of at least 2.618 × ATR leaves average deployment near 66%). The best filter set on development, RSI ≥ 70 with a bullish body of at least 60%, and range ≥ 1.0 × ATR, is the one carried into the combinations; its development lead over the baseline is not matched in a stable way across neighbouring filter values.

Exits

Month-end ATR stops with k = 1.0 (previous close minus one ATR) are the only exit change that improves both periods materially (10.0% development, 16.1% validation as an addition to the baseline exits). Intramonth stops monitored on daily bars were worse at tight k: whipsaw exits at k = 1.0 produce 722 round trips instead of 274, higher costs and a gap loss of about ₹17.8 lakh in the run; wider k (2.5-3.0) rarely trigger and change little. Stop-only exits with wide k (2.0-3.0) leave positions in falling trends too long (drawdown 39-68%) and trade very little (12-90 round trips).

Portfolio-wide exit and re-entry rule

As specified (T = 100, 75%) the rule never fires: with the first-breakout-in-six-months condition, the most qualifying signals in any month was 54 across the whole universe (43 among the 200 largest). Lower thresholds do fire (T = 20: 4 times; T = 10: 6 to 12 times) but the effect is inconsistent (CAGR from 6% to 12% across the ten settings) and not distinguishable from noise. Exits execute at the next month's open, so the gap between the month-end close and that open is borne by the account; it is inside the trade-level exit prices.

Sizing, concentration and capital utilisation

Position sizing mostly trades return for utilisation: pure risk-based sizing (stop distance from the SMA or Supertrend, clipped to 5-35%; 0.5-1.5% of equity at risk) deploys 46%-91% of capital across the settings, generally with lower drawdown and lower return than the equal-slot rules (chart); scaling the risk-sized positions up pro rata to use the idle cash (RISKFILL), capped per stock, brings deployment to 56%-93%. Equal-slot sizing with 50 slots deploys most capital at a lower drawdown than 10 slots (diversification), at the cost of more small positions. The per-stock cap rows (5-15%) and the sector cap rows (25%, 35%) are in the configuration table; sector classification covers only names in the current index list, so the sector cap cannot constrain delisted names. Portfolio drawdown controls (no new entries below −15% or −20% until the drawdown halves) kept the account out of the market for long stretches (average deployment 23%-36%) and gave up most of the return: not recommended.

Worked sizing examples for the ₹50 lakh account (finalist rules: RISKFILL, 1% risk, 20 names, 10% cap)

Method. (1) Equity E = cash + positions at the prior close. (2) Stop distance d = (signal close − max(SMA20, Supertrend line)) ÷ signal close, clipped to 5-35%. (3) Risk-based amount = 1% × E ÷ d, capped at 10% × E. (4) RISKFILL multiplies the risk-sized amounts by one factor so that the month's new positions together use the available cash, never exceeding the 10% cap. (5) Shares = floor(amount ÷ (open × (1 + slippage) × (1 + charges))), at the raw open price. (6) Rupee risk = cost × d. Examples taken from the finalist's actual signals:

Example reading: with E = ₹50,00,000, a stock whose stop reference is 12% below the signal close gets 1% × 50,00,000 ÷ 12% = ₹4.17 lakh; if only two signals appear and cash is plentiful, RISKFILL scales both up (up to the 10% cap = ₹5 lakh each), so about 20% of capital is deployed rather than about 17%. When many signals compete for few free slots, the liquidity ranking decides and the rest are counted as missed (278 of 562 signals in the finalist run at 2x).

6. Regimes, trades and concentration

Using the Nifty 200 against its 10-month average as the regime: in 169 bullish months the finalist earned 30.2% a year against 28.2% for the index (positive in 70% of months); in 65 bearish months -16.9% against -27.3% (49% positive). The finalist's regime filter holds back entries in bear regimes but does not exit open positions (the 'exit' variants are in the search table).

The best 10 names produced 78% of net trading P&L across 198 names traded; breakout strategies make money from a small share of large winners (win rate 45%). Net P&L by sector (names with a current classification):

7. Recommendation and limits

1. Do not paper-trade or fund the finalist. Its edge over the Nifty 200 is not distinguishable from selection among 211 configurations (reality check p = 0.70; fresh-start validation 13.0% against 14.0%).
2. If a monthly Bollinger breakout is used at all, the supported refinements are: the month-end ATR(14) stop at k = 1.0 as an additional exit, an index trend regime filter on entries, and risk-based sizing scaled to use idle cash with a 10% per-stock cap and 20 names. Drawdown controls and the intramonth stop at tight k are not supported.
3. The portfolio-wide exit as specified cannot trigger; a lower threshold needs its own pre-registered study.
4. These results are not independently validated: the same data pipeline, a point-in-time stand-in for the index, price-only returns, heuristic corporate-action handling, T+0 cash recycling, daily-bar stop assumptions (a session whose low touches the stop is assumed filled) and a sealed period that remains unopened.

Files: research/bbm/ (prep.py, sim.py, run_search.py, final.py, export_bbm.py, report_bbm.py, results/), data/research/panel_sf.parquet.

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.