EXP-016 — Separating relative and absolute ETF momentum

EXP-016 — Separating relative and absolute ETF momentum

• Status: conditional design complete · not run, waiting for product and distribution data gates
• Contract: common contract
• Hypothesis: combining cross-asset relative strength with a positive-trend filter improves after-cost performance versus holding the assets at equal weight.
• This design is a monthly long-only rule. It does not expect the returns of long-short futures time-series momentum research.

Product selection — fixed without looking at returns

Four asset classes: a domestic broad equity index, a US broad equity index (currency-exposed), domestic government bonds (medium term), spot gold. Leveraged, inverse, ETN, synthetic derivative-strategy and duplicate same-index products are excluded. Cash is held as a KRW balance at 0% interest rather than finding a separate fifth ETF.

• Candidate sources: the ETF listing at the time, tracked indexes, asset/currency-hedge classification and distribution data.
• The selection date is the last KRX trading day of 2018. In each asset class, among products with 252 trading days of history up to that day and 60-day trading-value data, the one with the largest median trading value is chosen. Ties by symbol code.
• If any class has no eligible product, data_hold. It is not arbitrarily replaced by a current large ETF or gold futures. If a shortened experiment with a later selection date is needed, register a separate ID and the reason for shortening.
• If a product is delisted or merged during operation after selection, the settlement known at the time is applied and that asset class goes to cash. It is not automatically linked to a successor ETF that performed well. Without a point-in-time candidate list, residual selection bias is flagged.
• The actual symbol codes are an output of the next preparation step under this rule; tickers not yet confirmed are not written here.

Data gates

1. Raw trading prices, splits/consolidations, trading halts, ex-distribution and payment dates and amounts, listing/product classification at the time.
2. Build the total-return signal index TRI from distributions confirmed up to each date. Record the data version to ensure that future revisions of the latest total-return index do not change past rankings.
3. Distinguish the signal's TRI calculation from the account's cash distributions. If total-return adjusted prices are used as fill prices, distributions are not added to the account again. Until the payment date, the entitlement to the distribution is distinguished from cash.
4. Domestically listed foreign-asset ETFs use KRW prices at KRX trading times. The same date's US close is not added to the Korean market-close signal. The currency effect is also part of this product's return.
5. Verify taxes, tax bases and account assumptions for domestic equity / other ETFs. If insufficient, only pre-tax exploration is possible; after_tax_advance is forbidden. Lack of data is not a reason to reject the strategy.

Monthly rule

• At the month-end KRX close, compute M12 = TRI_t / TRI_(month-end 12 months earlier) - 1.
• Target 95% of NAV in the one product ranked 1 on relative strength with M12>0; if no product is positive, cash. Ties by symbol code. A product without an exact 12-month month-end observation cannot be newly added that month.
• Sell at the next KRX trading day's open; if the proceeds are needed to buy the new asset, buy at the open of the following trading day. This follows the common accounting rule against using open-sale proceeds early. Buys within existing cash can happen on the first day.
• If the target product is unchanged, it is held without unnecessary monthly trading. The new target quantity is set once from NAV on the decision date with a 95% budget, and price changes during an execution delay do not change the target asset.
• Cash distributions accumulate in cash and are included in the budget at the next actual switch. Valuation returns and reinvestment are kept consistent.

Separating causes with a 2×2 comparison

Rebalancing of the equal-weight / absolute-filter accounts also runs with the same two-step sell/buy and costs. The 1st statistic is the daily log-return difference of dual − equal_weight. The filter's contribution is compared with the other two specifications under the same 12-month setting, and the best specification is not chosen as the base form after the fact.

Periods, matrix and verdict

• Returns in the calibration period used for product selection are not performance. Only two periods, history_A 2019~2022 and history_B 2023~2026-08-26, are evaluated as independent accounts. The 12-month warm-up uses data before the selection date.
• 4 specifications × 2 periods × 4 costs = 32 cells.
• The 6-month signal for dual only: 2 periods × 4 costs = 8 diagnostic cells. 40 cells in total; the cash comparison is separate.
• 1st cell: dual/12 months/history_B/proxy_slip2. The sample needs at least 36 completed months and at least 12 actual asset switches. If short, the period is not extended at will; hold.
• Research promotion: data requirements met, own after-cost net return >0, lower bound of the 1st CI >0, history_A excess point estimate >0, own net return and excess effect positive under both cost stresses, net P&L positive excluding the best year.
• Interpret total holding weight, MDD, turnover and the differences by relative/absolute factor together. The effect of simply losing less by staying in cash for long periods is not called a relative-rank effect.
• If raw prices, distributions and taxes cannot be verified, no operational promotion even with good numbers.

Validation and future outputs

Synthetic tests cover month-end holidays, the 12-month calculation, time lags, ex-distribution vs actual payment, keeping the same asset, all-negative → cash, delisting, trading halts, two-step switches and inability to buy integer quantities. A product selection manifest, TRI and cash dividend ledgers, monthly ranks/intent/actual weights, the effects of the two comparison factors, NAV, underwater curve, monthly heatmap, switching costs and sample gates will be recorded later. Data collection has not been run yet either.