15.8 Playbook 8: gold through the import-parity lens (GOLDBEES around the May 2026 duty change)

15.8.1Prerequisites and what this playbook teaches#

Read first (essential) M10 · 10.5 (import parity, the duty history and (10.15)–(10.16)), M10 · 10.6 (what drives gold, especially real yields), M11 · 11.5 and 11.11 (real rates and the rupee), M03 · 3.1.1 (how many trades distinguish edge from noise) and 3.9 (position sizing), M01 · 1.7–1.8 (costs and tax).

For depth: M10 · 10.12 (CFTC positioning in gold) and M13A · 13A.7 (half-lives of a mean-reverting premium).

What it teaches

1. How to turn the import-parity formula into a daily series and read an Indian gold ETF against it: which part of a rupee gold move came from London or New York, which from the rupee, and which from the Finance Ministry.
2. What a policy jump looks like in the data. The 13 May 2026 duty increase moved the rupee gold price by about 8.5% overnight, with no warning.
3. Why a relative-value signal (the ETF is cheap relative to parity) is a weak reason for an outright long position: the signal is worth about 2.5%, while gold can move more than that in a week.
4. How a baseline estimated across a regime change produces false signals.

The trade below is a paper trade that lost 1R. It is an educational reconstruction on public data, not a recommendation to buy, sell or hold gold or any fund.

Instrument Nippon India ETF Gold BeES (NSE symbol GOLDBEES) is an exchange-traded fund (ETF): a mutual-fund scheme whose units trade on the exchange like shares and whose assets are physical gold held by a custodian. It is the most liquid gold ETF on NSE and the tradeable proxy used here because MCX futures data are not freely available. Its units settle T+1 into a demat account like equity.

15.8.2The setup as testable rules#

All quantities are computed after the US close of day $t$, which is about 02:30 IST on day $t+1$, so they are known before NSE opens on $t+1$.

Parity From (10.15), with Yahoo's COMEX front-month gold close $P^{\$}_t$ (symbol GC=F, US dollars per troy ounce), Yahoo's USDINR close $X_t$ (INR=X) and the duty $d_t$ (BCD plus AIDC: 15% until 23 Jul 2024, 6% from 24 Jul 2024, 15% from 13 May 2026; M10 · 10.5.2), the duty-paid parity in rupees per gram is

Implied grams per unit With $C_t$ the GOLDBEES close, $u_t = C_t/\Pi_t$ is the quantity of duty-paid fine gold that one unit is priced at. It drifts down slowly because the fund deducts its expenses from the gold it holds, and it jumps when the domestic price moves to a premium or discount to parity (the term $q$ of (10.15)).

Premium or discount $\rho_t = u_t/\operatorname{median}(u_{t-249},\dots,u_t) - 1$, the deviation from the usual relation over the last 250 trading days, and $\bar\rho_t$ its 5-day mean. The averaging matters: COMEX closes about eleven hours after NSE, so the daily $\rho_t$ contains whatever gold did in the US session. Its standard deviation over 2015–2026 is 1.50%; that of $\bar\rho_t$ is 1.28%.

Rules The parameters are round numbers, not optimised; the base rates in 15.8.7 are nonetheless in-sample.

• Signal: $\bar\rho_t \le -2.5\%$ (the ETF is at least 2.5% cheap relative to its usual relation with parity) and the US 10-year real yield (FRED DFII10, the previous US business day's value) has not risen over the last 20 trading days. The second condition is the opportunity-cost driver of M10 · 10.6.2: do not buy gold into rising real yields.
• Entry: market order at the next NSE open. Size with (3.23) at 1% of ₹10,00,000 using the pre-open price.
• Stop: a resting stop-loss order at the signal-day close minus $2.5\times$ the 14-day average absolute close-to-close change (a close-based ATR; Yahoo's GOLDBEES open, high and low fields contain bad prints, so the volatility estimate uses closes). If the market opens below the stop, the fill is the open.
• Exit: at the next open after $\bar\rho_t \ge -0.5\%$ (the discount has closed), or after 40 trading days, whichever comes first.
• Invalidation: any change in the gold duty or a CBIC tariff-value change large enough to move parity by more than 1%: exit at the next open, because the baseline no longer describes the market.

The thesis is a convergence thesis: the ETF should recover its usual relation with parity. The position, however, is an outright long. Only the relative part of the move is predicted; the gold price itself is not.

15.8.3The scan and the episode#

Data Yahoo daily bars for GOLDBEES.NS, GC=F and INR=X, and FRED DFII10, from 1 Jan 2015 to 6 Nov 2026 (2,920 GOLDBEES trading days after removing one mis-scaled row on 19 Dec 2019). Fetched 6 Nov 2026.

The duty change as an event On 12 May 2026 GOLDBEES closed at ₹124.61. The duty increase from 6% to 15% was reported as effective immediately on 13 May (M10 · 10.5.2). GOLDBEES opened at ₹131.10 (+5.21%), traded as high as ₹134.40 and closed at ₹131.74 (+5.72%). Decompose the move with (10.16):

Component (12 → 13 May 2026, log)	Value
COMEX gold, $4,686.70 → $4,706.70	+0.43%
USDINR, 95.3914 → 95.6330	+0.25%
Duty, $\ln(1.15/1.06)$	+8.15%
Parity, ₹15,236.08 → ₹16,642.30 per gram	+8.83%
GOLDBEES	+5.56%
Change in premium/discount ($\ln u$)	−3.26%

The ETF captured about two-thirds of the parity jump. The domestic market moved from roughly fair to a discount of about 3% to parity on the day, and stayed at a discount: the median of $u_t$ from 24 Jul 2024 to 12 May 2026 was 0.008279 g per unit, and from 13 May to 6 Nov 2026 it was 0.008113, 2.0% lower. This is the pattern M10 · 10.5 predicts: a duty increase depresses domestic demand and encourages smuggling, so the local price settles below full parity. The July 2024 duty cut shows the mirror image. From 22 to 24 Jul 2024 parity fell 7.30% (log) while GOLDBEES fell only 4.08% (log), so the ETF moved to a premium of about 3.2%.

Could anyone have positioned for 13 May? No. The duty is a policy variable, announced without notice. A long holder of ₹3 lakh of GOLDBEES on 12 May (2,407 units) gained ₹17,162 overnight; a holder short of rupee gold through MCX would have lost the full jump. The lesson is about jump risk: an Indian gold position carries a policy term that COMEX gold does not, and it can only be sized for, not forecast.

What the rule saw in 2026 The 250-day baseline did not know about the regime change, so after 13 May $\bar\rho_t$ was persistently negative. It crossed −2.5% on these days:

Date	Close (₹)	$\bar\rho_t$	20-day change in DFII10 (pp)	Rule
18 May 2026	129.70	−2.61%	+0.17	blocked: real yields rising
19 May 2026	130.64	−2.71%	+0.23	blocked
19 Aug 2026	125.73	−2.51%	+0.04	blocked
20 Aug 2026	128.88	−2.67%	−0.04	signal

The episode is therefore the 20 Aug 2026 signal, chosen by the rule. The May signals were filtered out; what they would have done is in 15.8.7.

15.8.4The plan before entry#

Written on the evening of 20 Aug 2026 after the US close, with the inputs then available.

Inputs at the signal GOLDBEES close ₹128.88; COMEX gold $4,571.40; USDINR 95.4960; parity ₹16,140.74 per gram; implied grams per unit 0.0079848 against a 250-day median of 0.0081941, so $\rho_t = -2.55\%$ and $\bar\rho_t = -2.67\%$; DFII10 2.35%, down 0.04 pp over 20 days; close-based ATR ₹1.145.

Thesis The ETF is about 2.7% cheap relative to its usual relation with import parity, and the real-yield driver is not adverse. If the discount closes while gold and the rupee are unchanged, the ETF gains about 2.7%.

Trigger Buy at the 21 Aug open.

Stop $128.88 - 2.5 \times 1.145 = \text{₹}126.02$, placed as a stop-loss order immediately after the fill.

Exit Next open after $\bar\rho_t \ge -0.5\%$; time stop 40 trading days; invalidation on a duty change.

Size (M03 (3.23)) GOLDBEES opened at ₹130.14, 1.0% above the signal close, so the risk per unit is $130.14 - 126.02 = \text{₹}4.12$ and

a position of ₹3,15,849.78 (31.6% of capital) with a loss at the stop of ₹9,999.24. Sizing from the signal close instead (risk ₹2.86) would have given 3,496 units and a 1.44R loss at the same stop. The size must be computed from the price actually paid.

Costs (M01 · 1.7, rates as of 6 Nov 2026) Brokerage for delivery is set to zero as an illustrative parameter. No STT: securities transaction tax under Chapter VII of the Finance (No. 2) Act, 2004 is levied on equity shares, equity derivatives and units of equity-oriented funds; a gold ETF holds gold, not equity, so on our reading of the statute its units attract no STT (not confirmed against an exchange circular; check before relying on it). Exchange charge 0.00307% of turnover and stamp duty 0.015% on the buy are assumed to follow the equity-delivery rows of the M01 table; SEBI fee ₹10 per crore; GST 18% on exchange charge plus SEBI fee; DP charge ₹15.34 on the sale. For this position the round trip costs about ₹86, 2.7 bp of the purchase value, or 0.009R. Add one tick (₹0.01) of slippage on each side: ₹49.

Expected R and break-even win rate From the rule's 13 trades from 2016 that had closed before 20 Aug 2026: win rate 46.2%, average win 1.10R, average loss 0.87R, mean +0.041R, standard deviation 1.46R. The break-even win rate is $0.87/(1.10+0.87) = 44.1\%$. The expectancy is indistinguishable from zero ($t = 0.10$). The honest plan says so: this is a probationary paper trade under the protocol of M14 · 14.10.3, taken to see whether live behaviour matches the backtest distribution, not because an edge has been demonstrated.

15.8.5Day-by-day management#

Prices are Yahoo daily bars. Parity uses the same day's COMEX and USDINR closes; DFII10 is the previous US business day's value.

Date	GOLDBEES O / L / C (₹)	COMEX $	USDINR	Parity ₹/g	$\bar\rho$	DFII10 %	Action and notes
Thu 20 Aug	127.83 / 127.83 / 128.88	4,571.40	95.4960	16,140.7	−2.67%	2.35	Signal after US close. Plan written.
Fri 21 Aug	130.14 / 129.82 / 131.32	4,680.60	95.7722	16,574.1	−2.84%	2.35	Bought 2,427 at 130.14 (open). Stop order at 126.02.
Mon 24 Aug	132.86 / 132.53 / 133.02	4,697.80	95.7003	16,622.5	−2.85%	2.40	Open profit 0.70R. Discount unchanged: the ETF rose with gold, not relative to it.
Tue 25 Aug	133.02 / 132.28 / 132.78	4,694.50	95.7231	16,614.8	−3.06%	2.38	Discount at its widest. Hold.
Wed 26 Aug	132.31 / 132.25 / 132.37	4,653.30	93.5461	16,094.4	−2.03%	2.32	The USDINR print of 93.55 reverses the next day and is probably a bad quote; the 5-day mean damps it.
Thu 27 Aug	131.24 / 128.15 / 130.05	4,664.00	95.4303	16,456.4	−2.21%	2.34	Low 128.15, 2.1 above the stop.
Fri 28 Aug	129.65 / 129.13 / 130.85	4,529.90	95.4704	15,989.9	−1.56%	2.34	COMEX −2.9%; the discount narrows because parity fell more than the ETF.
Mon 31 Aug	127.79 / 126.08 / 127.32	4,481.50	95.3803	15,804.1	−1.41%	2.42	Low 126.08, six paise above the stop. Rules forbid widening it (M04A · 4A.10).
Tue 1 Sep	127.19 / 125.02 / 125.34	4,396.40	95.1144	15,460.8	−1.12%	2.44	Stop triggered intraday; filled at 126.02 (assumed, stop-loss market order).

15.8.6Exit and result#

Item	₹
Bought 2,427 at 130.14	3,15,849.78
Sold 2,427 at 126.02	3,05,850.54
Gross P&L	−9,999.24
Exchange 19.09, SEBI 0.62, stamp 47.38, GST 3.55, DP 15.34, STT 0	−85.97
Net P&L	−10,085.21 (−1.009R)
With one tick of slippage on each side	−10,133.75 (−1.013R)

Tax Held 8 trading days, so the result is a short-term capital loss. The Finance (No. 2) Act, 2024 is reported to have narrowed the definition of a "specified mutual fund" (whose gains are always taxed at the slab rate) to funds investing mostly in debt. The common reading is that gold ETF units therefore became ordinary capital assets: held for more than 12 months, long-term and taxed at 12.5% without indexation; otherwise short-term and taxed at the slab rate. Readings differ on how the change applies to units bought before and sold after 1 Apr 2025. We have not verified this against the text of the Act or of the Income-tax Act, 2025; treat it as unverified and check it with M00 · 0.7 and a chartered accountant. A short-term capital loss can be set off against capital gains of the year and carried forward 8 years (M01 · 1.8); its after-tax value depends on what it is set off against, which M00 · 0.7 discusses. The pre-tax result is −₹10,085.

Where the loss came from From the signal close (20 Aug) to the exit day (1 Sep), in logs: COMEX gold −3.90%, USDINR −0.40%, so parity −4.30%; GOLDBEES −2.79%. The ETF outperformed parity by 1.52 percentage points: the relative thesis was about half right. The outright gold move, which the rule does not predict, was nearly three times as large and of the opposite sign. The real-yield filter did not help: DFII10 rose from 2.35% to 2.44% during the trade.

15.8.7Post-mortem#

Was the process followed? Yes. Size from the fill, stop placed at once, no widening on 31 Aug when the low came within six paise. The loss was exactly the planned 1R plus costs.

Base rates for the rule, 2016–2026

Version	$N$	Win rate	Mean R	SD R	Total R
Rule as specified (with the real-yield filter)	14	42.9%	−0.034	1.43	−0.47
Trades known at the 20 Aug signal	13	46.2%	+0.041	1.46	+0.53
Without the real-yield filter	14	50.0%	+0.087	1.52	+1.22

The median holding period is about 10 trading days. Without the filter, the rule would have bought on 19 May at ₹130.68 and been stopped by a gap on 8 Jun at ₹124.98 (−1.23R), and would also have taken the August trade. Neither version is distinguishable from zero. Do not now drop or keep the filter because of these numbers: with 14 trades the difference between the rows is noise, and choosing the better row after the fact is the selection error of M03 · 3.1.1.

How many trades would be needed? With the pre-trade estimates $E = 0.041$R and $s = 1.46$R, M03 (3.2a) gives $N = (2s/E)^2 \approx 5{,}100$ trades for $t = 2$ in expectation and (3.2b) about 10,400 for 80% power. At about 1.3 signals a year this edge, if it exists, can never be established from the trading record. The rule is a hypothesis, not a strategy.

What was wrong with the design

1. The baseline crossed a regime change. On 20 Aug, 180 of the 250 days behind the median of implied grams per unit came from the 6%-duty regime and only 70 from after the hike. Measured against the median since 13 May (0.008092), $\bar\rho$ on 20 Aug was −1.44%, not −2.67%: no signal. The invalidation rule exited trades on a duty change but did not reset the baseline afterwards. M10 · 10.5's practitioner note says that discounts persist after duty increases; the rule should have encoded it. A fix (reset the baseline at each duty change and require 60 days of data before trading) must be tested on the whole history, including the 2024 cut, before it is adopted, and that history contains only two duty changes.
2. Most of the signal is timing noise. The daily $\rho_t$ has a standard deviation of 1.50% and the threshold is 2.5%. A large part of a typical "discount" is COMEX moving after NSE closes; the next NSE open absorbs it. The 21 Aug open was 1.0% above the signal close for this reason, and the entry paid for it.
3. A relative-value signal in an outright position. The convergence earned +1.5% while the gold move cost −4.3%. To trade only the relative part one would need to short COMEX gold and buy dollars against rupees in proportion, which a retail account in India cannot do cheaply. Without the hedge the parity reading is better used as a valuation check on an allocation decision (do not add to gold when the ETF is at a large premium to parity) than as a trading signal.

What the episode does teach The import-parity decomposition explains every rupee of the 13 May jump and of the August loss, and that is its real use: knowing whether a move in Indian gold came from London, the rupee or the government.

15.8.8Do it in the simulator#

The simulator (Simulator → New order) works on end-of-day data: GOLDBEES.NS orders fill against Yahoo daily bars; a market order fills at the day's close, or at the next open if your account's market-fill setting is "open"; stop orders become eligible from the next open and fill at the open if the price gaps through them. It charges the M01 equity-delivery cost stack, which includes 0.1% STT on each side. If, as 15.8.4 reads the statute, gold ETF units pay no STT, that overstates the cost by about ₹622 on a round trip the size of the one above. Note the difference in your journal.

1. Scan. Using the lab below or your own script, compute $\Pi_t$, $u_t$, $\bar\rho_t$ and the 20-day change in DFII10 for the latest available day. Use the post-13-May baseline as well as the 250-day one, and record both.
2. Decide. If the rule as written in 15.8.2 signals, set the account's market-fill to "open". Under Simulator → New order enter GOLDBEES.NS, buy, quantity from (3.23) at 1% of your simulator equity, using the latest close plus 1% as the assumed fill. Set the stop loss at close minus 2.5 close-ATRs and leave the target empty, because the exit is a rule condition, not a price. Write the thesis in the parity terms of 15.8.4. Check in Preview that the risk shown is at most 1% of equity.
3. If there is no signal, place no order. Record the scan values in the journal; a no-trade day is a result.
4. Manage. Each evening recompute $\bar\rho_t$. If it reaches −0.5%, close at the next open with a market order. After 40 trading days, close. Never move the stop down.
5. Review. At exit, decompose the trade's return into the gold, FX and premium terms as in 15.8.6, and record the R-multiple with and without the simulator's STT.

See it live: current prices of the listed gold ETFs at https://oicharts.com/etfs/gold .