Agent Adda · Cross-Market Desk Note 17–18 Sep 2026

Copper, Gold & Crude: What the Backtests Actually Show

A side-by-side read of three independent strategy studies run over the past two weeks — a 7-year Supertrend+RSI swing system on copper, and MCX Crude Oil / Gold Mini intraday systems tested this week. Same discipline throughout: train/test split, cost-adjusted, and every result graded on the out-of-sample slice, not the fit.

MCX Crude Oil
Bollinger Squeeze holds up
PF 2.19 full-sample, PF 2.20 out-of-sample — the one setup where test confirms train.
MCX Gold Mini
Only trend-following survives
Supertrend Flip is the sole positive-PF system out-of-sample; breakout variants bled money both halves.
Copper
Supertrend + RSI filter wins
7yr backtest: +35.7% return, PF 1.29×, −29.1% DD — beats plain Supertrend on drawdown for the same return.

MCX Crude Oil — 30-min intraday

8 contracts, 7 rollovers
17 Mar – 17 Sep 2026

Continuous 30-minute series stitched across eight MCX contract rolls, tested with a 1.5×ATR stop, 16-bar max hold, forced flat at session close, and 2 bps/leg costs on a 60/40 train/test split. Bollinger Squeeze is the standout: it is the only strategy whose out-of-sample profit factor (2.20) matches its full-sample number (2.19), with the tightest full-sample drawdown per lot of the five systems tested.

₹10,538 ₹8,872 ₹6,979₹8,033₹9,088₹10,142₹11,196 03-1705-0406-1808-0309-17
MCX Crude Oil — continuous daily close, 17 Mar–17 Sep 2026 · +18.8% over the window

Strategy leaderboard — ranked by test profit factor

StrategyTrades (full)Win % (full)PF (full)Net / lot (full)Max DD / lotPF (test)Net / lot (test)
Bollinger Squeeze6949.3%2.19₹311,497₹−45,0132.20₹111,425
Darvas + Supertrend (MTF)4839.6%1.56₹138,573₹−72,6962.21₹76,228
Darvas Breakout8246.3%1.69₹261,505₹−61,7051.50₹72,386
Darvas + Supertrend8245.1%1.68₹257,709₹−61,7051.42₹61,089
Supertrend Flip13240.9%1.06₹40,554₹−150,7661.86₹140,737
Read: every one of the five systems finished the out-of-sample slice net-positive per lot — unusual breadth, but treat it as a favorable regime (broad trending March–September tape) as much as a strategy edge. Bollinger Squeeze is the pick precisely because train and test agree, not because it has the biggest number.

MCX Gold Mini (GOLDM) — 30-min intraday

7 contracts, 6 rollovers
9 Mar – 7 Sep 2026

Same engine, same cost and stop assumptions as crude. The result set is far less forgiving: four of five strategies lose money out-of-sample. Supertrend Flip is the exception and the only one worth carrying forward — test profit factor actually improves to 2.04 from a full-sample 1.30, meaning its edge concentrated in the more recent half of the data rather than decaying.

₹143,338 ₹160,299 ₹130,385₹139,087₹147,790₹156,492₹165,194 03-0904-2306-0807-2309-07
MCX Gold Mini — continuous daily close, 9 Mar–7 Sep 2026 · −10.6% over the window

Strategy leaderboard — ranked by test profit factor

StrategyTrades (full)Win % (full)PF (full)Net / lot (full)Max DD / lotPF (test)Net / lot (test)
Supertrend Flip12245.9%1.30₹95,286₹−69,0512.04₹102,576
Darvas + Supertrend (MTF)4833.3%1.11₹20,234₹−71,6071.06₹4,012
Bollinger Squeeze7045.7%1.06₹11,640₹−60,6260.80₹−17,525
Darvas + Supertrend10233.3%0.96₹−13,519₹−109,6090.80₹−34,874
Darvas Breakout10233.3%0.93₹−26,362₹−114,1350.80₹−34,874
Data gap: the GOLDM series ends 7 Sep 2026 (cache-only, no live refresh in this run) — ten days stale versus the crude and copper studies. Directionally usable, but re-run before sizing anything off it.

Copper — Supertrend + RSI swing system

COMEX HG=F, daily, 1,762 bars
7yr: Sep 2019 – Sep 2026

The core copper study: three swing strategies backtested on 7 years of daily COMEX HG=F bars, $100k notional, 1 contract = 25,000 lbs, slippage not modelled. Supertrend(10,3) + RSI<68 filter is the winner — it matches pure Supertrend on return while cutting max drawdown by nearly 2 points, by refusing entries into overbought extensions exactly like today's.

Strategy leaderboard — 7-year backtest

StrategyTradesWin %Total ReturnSharpeMax DDProfit Factor
Supertrend(10,3) + RSI<68 filter10423.1%+35.7%0.33−29.1%1.29×
Supertrend(10,3) only11023.6%+34.7%−30.9%
Darvas Box + Supertrend3638.9%+14.0%−26.7%
The 23% win rate is structural, not a flaw — copper trends in long runs with frequent small whipsaws. Winners average $6,638 vs. losers $1,545, and PF 1.29× means every $1 lost on losing trades earns $1.29 back on winners. The RSI filter's entire job is blocking chase-entries into extensions like the RSI 70.3 print on 8 Sep — it improved Sharpe without giving up return.

Entry / exit rules (Supertrend + RSI filter)

Entry trigger
ST(10,3) flips bullish, daily close, AND RSI < 68
Preferred entry
1–3% pullback from high, RSI cools to 60–65
Exit trigger
ST(10,3) flips bearish, daily close → exit next open
Overbought trim
RSI > 78 + bearish divergence → cut 30%

Where the levels stood at publish (8 Sep 2026) vs. today

At the 8 Sep publish, COMEX closed at an all-time high of $6.849 with RSI 70.3 — the filter explicitly blocked a fresh entry and called for a pullback to $6.60–6.70 (MCX ₹1,360–1,380) with a $6.45 (₹1,330) hard stop. That call played out: the September MCX contract bottomed at ₹1,350.30, never touched the ₹1,330 stop, and rebounded to ₹1,395.50 by 17 Sep evening. COMEX itself dipped to $6.33 on 14 Sep — briefly invalidating the COMEX-only version of the trade — before recovering to $6.647 with RSI 59.2 by 17 Sep, still below its own Supertrend line, so the mechanical system has not re-triggered a fresh entry.

Original entry zone
₹1,360–1,380
Original hard stop
₹1,330
Target 1 / Target 2
₹1,484 / ₹1,608
Updated swing-low invalidation
₹1,349
Updated reclaim zone
₹1,378–1,385
Updated de-risk zone
₹1,412–1,431
LME stocks rose to 254,150t (from 237,725t on 8 Sep) while the Yangshan warrant premium firmed to $121/t — the earlier tariff-driven squeeze premium has cooled but physical demand signals have improved. September MCX contract enters tender on 28 Sep: exit or roll before then.

Follow-up quant check — combination sweep & long-horizon momentum

A follow-on study this week tried to sharpen the swing system by layering VWAP / Volume Profile / VCP filters on the same MCX daily setup, walk-forward split into a 2019–2023 train window, a 2023–2025 test window, and an Apr 2025–Sep 2026 holdout — the true out-of-sample slice charted below. Every filtered combination still came back under-sampled on the holdout (4–7 trades in ~17 months), so treat the ranking as a lead, not a signal; what the chart does show clearly is that the unfiltered baseline system round-tripped its holdout gains back to a loss, while the VWAP variant held onto a positive edge and VCP-gated entries sat out and drifted flat-to-down on look-ahead-free equity.

VWAP ₹101,356Baseline ₹94,455VCP ₹93,499 ₹92,455₹97,456₹102,456₹107,456₹112,456 04-2108-2701-0605-0809-16
Baseline (no filter) VWAP filter VCP filter
$100k notional · holdout Apr 2025–Sep 2026 · dashed line = starting capital

The one filter result that cleared both a live walk-forward test and a 2002–2018 historical backcast on its own (not layered onto the combination sweep) was 252-day time-series momentum: +27.4% return / PF 1.31 on the 17-year backcast (58.8% of years positive), and +6.26% / PF 1.21 on its own live out-of-sample window — a slower, trend-confirming cousin of the Supertrend system above rather than a replacement for it.

Methodology

Crude & Gold engine
30-min continuous futures, stitched across contract rolls (Economic Times public MCX chart feed). 1.5×ATR stop, 16-bar max hold, flat at session end, 2 bps/leg cost, 60/40 train/test split.
Copper engine
7-year daily COMEX HG=F backtest (Sep 2019–Sep 2026, 1,762 bars), $100k notional, 25,000 lbs/contract, slippage not modelled. Follow-up sweep tests VWAP/Volume Profile/VCP filters on MCX daily bars, plus a standalone walk-forward spike test with a 2002–2018 historical backcast for the TSMOM family.
Reading "test"
Out-of-sample slice only — the number that matters. A strategy whose test PF holds up near its full-sample PF is doing real work; one where test collapses is fit to the train window.
Under-sampled
Fewer trades than needed for statistical confidence (rule of thumb: <30 on the test slice). Treat as a lead to keep testing, not a signal to trade.