Bot A Admin BT · 5yr Python BT · TradingView

Bot A — TradingView backtest, 2016-10 → 2026-10

196 NSE stock futures · 4H SuperTrend · LONG only · validated 2026-10-06 · RESEARCH — NOT DEPLOYED

Headline — risk cap ₹25,000 vs ₹35,000

2,124 / 3,107

trades 25k / 35k

56.1% / 54.9%

win rate

1.539 / 1.385

profit factor

₹63.4L / ₹85.7L

net

₹2,984 / ₹2,759

net per trade

₹0.181 / ₹0.133

net per ₹1 risked

₹10.3L / ₹20.4L

max drawdown

6.13 / 4.21

net / maxDD

₹35,000 was adopted — and not because it earns more. On the pooled 11 years ₹25,000 wins every quality measure. The reason to override that is a structural break at 2024: NSE raised the minimum F&O contract value to ₹15 lakh in Nov 2024, lifting lot sizes universe-wide. Since risk_per_lot = stop% × price × lot, a lot became 2–3× more expensive in risk terms, and a cap sized for ₹5–10 lakh contracts can no longer reach the current universe.

The deciding split

period	25k trades	25k net/trade	35k trades	35k net/trade	25k as % of 35k
2016–2023	1,818	₹3,158	2,266	₹2,845	79%
2024–2026	306	₹1,954	841	₹2,526	36%

Both halves of the case for ₹25,000 reverse. The cap used to drop 21% of signals and keep the better ones. It now drops 64% and keeps the worse ones.

Year by year

year	25k trd	25k net	25k /trd	35k trd	35k net	35k /trd	35k − 25k
2016	53	−68,525	−1,202	63	−174,286	−2,601	−105,761
2017	217	955,560	4,404	248	1,127,042	4,545	+171,482
2018	238	−40,733	−171	283	105,374	372	+146,107
2019	266	414,970	1,560	304	374,066	1,230	−40,904
2020	248	804,417	3,244	302	1,001,781	3,317	+197,364
2021	218	714,631	3,278	290	682,872	2,355	−31,759
2022	272	777,783	2,859	391	627,283	1,604	−150,500
2023	306	2,145,796	7,012	385	2,664,885	6,922	+519,089
2024	76	513,469	6,756	222	995,105	4,482	+481,636
2025	163	431,839	1,620	416	1,987,305	4,777	+1,555,466
2026	67	−347,528	−5,187	203	−858,026	−4,227	−510,498
TOTAL	2,124	6,338,944	2,984	3,107	8,570,663	2,759	+2,231,719

The 11 year rows sum to the batch totals within ₹2 of display rounding — the year table and the summary are internally consistent.

Per batch (5 × Pine scripts — the engine caps `request.*()` at 40 calls)

batch	syms	25k trd	25k WR	25k PF	25k net	35k trd	35k WR	35k PF	35k net
B1	40	407	53.6%	1.35	917,066	613	54.5%	1.39	1,840,267
B2	40	562	56.2%	1.60	1,868,012	793	53.6%	1.30	1,746,236
B3	40	321	55.5%	1.58	1,017,170	545	55.0%	1.45	1,858,776
B4	40	402	53.2%	1.40	909,782	566	52.8%	1.31	1,227,966
B5	36	432	61.6%	1.81	1,626,914	590	59.2%	1.52	1,897,418

Defect found and fixed during this study — B5 phantom trades. B5 held 36 real symbols but made 40 runSym() calls; slots 37–40 were left as input.symbol(""), and Pine resolves an empty symbol string to the chart symbol (ignore_invalid_symbol does not help — empty is valid). B5 was trading four phantom copies of whatever chart it sat on.

How it was caught: the same engine was run twice at ₹25,000 on two different chart symbols. B1–B4 reproduced to the rupee; B5 did not (600 trades / PF 1.48 → 636 / PF 1.75). Only B5 had empty slots. Guard applied: str.length(s37) > 0 ? c_37 : na.

The published 196-symbol headline was depressed, not inflated, by its phantoms — that run sat on BAJFINANCE, whose four copies lost ₹124,636. Corrected: 2,124 trades / PF 1.539 / ₹6,338,944 (previously 2,292 / 1.48 / ₹6,214,309).

Live config parity — what live was changed to match

dial	TV backtest	live before	live after	status
multiplier	2.5	0.5	2.5	matched
min_adx	20.0	25.0	20.0	matched
max_hold_days	10	5	10	matched
rr	2.0	2.0	2.0	already
min_atr_pct	0.1	0.10	0.10	already
atr_period	10	10	10	already
direction	LONG	LONG	LONG	already
risk / trade	₹15,000	₹15,000	₹15,000	already
daily-ST filter	none	off	off	already
NIFTY regime	none	off	off	already
MAX_RISK_PER_LOT	35,000	pinned 15,000	35,000	invariant broken
stop-distance filter	none	on, 0.6%	off	unmeasured
max_capital_inr	none	₹300,000	₹1,000,000	matched
max_concurrent	none	4	0 = unlimited	live risk
universe	196	90	196	196/196 on Dhan
bar source	futures 4H	equity 4H	unresolved	BLOCKED

Three risks accepted on the record.

The ₹35,000 cap breaks a deliberate safety invariant. Sizing is contracts = max(1, floor(risk / risk_per_lot)), so the 1-lot floor means a passing signal is taken even when one lot risks more than the target. Measured in this backtest: average risk/trade ₹20,696 — 38% above the ₹15,000 nominal, worst case ₹35,000 = 2.33×. Plan capital on ₹20,696, not ₹15,000.
Unlimited concurrency makes broker margin the only limit. Entries do reject cleanly when capital runs out — but if an entry fills and the bracket's SL leg is then rejected for margin, the position is naked. That is the most repeated failure in this codebase (BUG-011/018/020/032/034). Harmless while draft mode is on; must be re-evaluated before going live.
The stop-distance filter was switched off although it was validated. The 5.5-year sweep found it helped in every RR/ADX combination — but at mult 0.5, where bands sit close to price. At mult 2.5 the median stop is 4.74% against a 0.6% floor, so it should rarely bind. This was not measured.

What this backtest does not model

No concurrency cap. 2,124 trades assume unlimited simultaneous positions. Live now runs unlimited too, so margin is the binding constraint instead — a constraint the backtest never saw.
No margin gate. Real NRML margin is 17.7%–50.2% per symbol and is not applied here.
Max drawdown — MEASURED 2026-10-06. ₹1,033,528 at the ₹25,000 cap vs ₹2,036,908 at ₹35,000: +97% drawdown for +35% net, net/maxDD 6.13 → 4.21, and ₹25,000 wins 4 of 5 batches on risk-adjusted return. 🚨 Against ~₹13.3L of capital the ₹35,000 figure is ~153% of the account. Caveat: batch drawdowns cannot be summed exactly — the sum is an upper bound, the worst single batch (₹602,003 / ₹278,574) a lower bound, and for a LONG-only book the truth sits nearer the sum. The book is also uncapped, so live margin will cut both net and drawdown by an unknown amount.
One static lot size per symbol for all 10 years. NSE revises lots as prices rise, so early-year rupee figures are approximate. Trade counts and win rates are unaffected, and the cap decision rests on 2024–26 where current lots are roughly right.
Bar-level exit timing. Prices are exact; the moment inside the bar is not.
2026 loses under both caps (−₹347,528 / −₹858,026) and 2023 alone is 34% of all profit. Neither is addressed by the cap choice.

Evidence & source documents

.claude/evidence/bot_a_tv_backtest_evidence_2026_10_06.md — full audit trail: method and platform limits, three-way engine verification, every sweep table behind the config choice, the risk-cap study, the B5 defect with its proof, and the Pine bugs found and fixed.
.claude/guides/bot_a_tv_validated_config_2026_10_06.md — the validated config, IS/OOS, by-year results, the deployment change set and the three accepted risks.
.claude/guides/bot_a_real_futures_bt_2026_10_02.md — the independent Python study on real futures prices with lot sizing, historical lots, a real NRML margin gate and a measured drawdown. It reproduces live's loss at live's old dials (−₹258,164 / WR 31.3% vs live's −₹213,745 / WR 31.0%), which is what makes its other outputs worth reading.
pine/cap_B1..B5.pine — the five batch scripts behind every number on this page.

Method, in one line: both risk caps run side by side inside one script, so they see identical signals, bars and lot sizes — the only difference is the cap test.