Strategies

KOSPI Large-Cap Idiosyncratic Down-Shock Rebound: buy a Korean large cap at the next Seoul open after a session in which its return fell >=2.5 sigma below its beta-adjusted peer basket, fixed 5-session clock, 6% per leg, max 3 new legs per date / 6 concurrent (KREQ 1-DAY, long/flat, unlevered, 33 common shares)

Outcome: Optimizing

KospiLargeCapIdioDownShockRebound

Sample is adequate for optimization: 504 closed long legs on 382 distinct entry dates in the development period (2018-07 to 2025-01). That supports trade-count-sized walk-forward windows well above the 20-trade target.

Strategy report

Hypothesis

Original Research Lead hypothesis, not seeded. It applies the peer-relative down-shock template of lesson 266 to a venue the factory has barely used: KREQ has 1 experiment out of 5,010. The template has been promoted on Bybit (BTC residual), Hyperliquid (BTC residual) and ASX gold miners (peer-basket residual). Korea is a fresh market where the mechanism should be strong: retail investors make up a large share of KOSPI turnover, daily limits are +/-30%, and forced selling (margin calls on retail credit loans, fund redemptions) produces price-pressure drops that are not news about value. This is a new market for the template, not a transfer of any earlier result (L228). MECHANISM: when one large cap falls far more than the rest of the market on a given day, after adjusting for its beta, part of the drop is non-informational liquidity demand (retail stop-outs, margin calls, block sales, an index or ETF reweight). Liquidity providers who absorb it are paid by a partial reversal over the next week (the short-horizon reversal literature: Lehmann 1990, Jegadeesh 1990; Nagel 2012 frames it as the return to liquidity provision). Measuring the shock against peers removes market-wide moves, which are informational and do not revert in the same way. UNIVERSE (pre-registered, 33 KRX common shares, Yahoo daily bars from 2018-01 in the catalog): 005930 Samsung Elec, 000660 SK hynix, 402340 SK Square, 005380 Hyundai Motor, 009150 Samsung Electro-Mech, 373220 LG Energy Solution, 207940 Samsung Biologics, 028260 Samsung C&T, 105560 KB Financial, 012450 Hanwha Aerospace, 032830 Samsung Life, 000270 Kia, 055550 Shinhan, 329180 HD Hyundai Heavy, 034020 Doosan Enerbility, 068270 Celltrion, 012330 Hyundai Mobis, 006400 Samsung SDI, 066570 LG Electronics, 086790 Hana Financial, 035420 NAVER, 034730 SK Inc, 010120 LS Electric, 000810 Samsung Fire, 010130 Korea Zinc, 298040 Hyosung Heavy, 267260 HD Hyundai Electric, 042660 Hanwha Ocean, 009540 HD KSOE, 005490 POSCO Holdings, 316140 Woori Financial, 096770 SK Innovation, 017670 SK Telecom. All preferred-share lines (005935, 005385, 005387, 005389, 009155, 02826K, 03473K) are excluded on purpose. A preferred moves with its common share, so including them would put near-duplicates in the peer basket and fake shocks into the signal. Point in time: a name is eligible only once it has 60 completed sessions of bars, so later listings (LGES 2022-01, SK Square 2021-11, HD Hyundai Heavy 2021-09) enter automatically and never before they traded. SURVIVORSHIP CAVEAT: there is no point-in-time KOSPI-200 membership file, so the panel is drawn from today's large caps. That biases the basket's level return upward; SK hynix, the HD Hyundai names, Hanwha Aerospace and Hyosung Heavy were the 2024-26 winners. The claim under test is therefore the RELATIVE claim: the 5-session forward return after a shock beats the peer basket's 5-session return over the same days. The analyst should read the measured excess, not only the raw return. PRE-REGISTERED MEASUREMENT (frozen rule, catalog daily bars, gross, entry at the next open, exit at the open after 5 sessions). These figures document firing rate and economics; they were not used to accept the rule. 706 events on 493 distinct entry dates between 2018 and 2026-10. Mean raw forward return +1.55% (t = 4.9), median +0.80%, win rate 56%, PF 1.78. Mean excess over the peer basket on the same days +0.70%. Raw mean by year: 2018 +1.34%, 2019 +0.41%, 2020 +2.29%, 2021 +0.73%, 2022 -0.37%, 2023 +2.16%, 2024 +3.10%, 2025 +1.49%, 2026 YTD +2.96%. Excess by year: +1.47, +0.44, +0.44, +0.19, -0.20, +1.21, +1.67, +0.07, +1.39 (%). Neighbouring settings are smooth, with no cliff: 2.0 sigma / 5 sessions +1.10% raw (+0.40% excess, 866 dates); 3.0 sigma / 5 sessions +1.51% (+0.68%); 2.5 sigma / 3 sessions +0.89% (+0.49%). Concentration: 367 of the 493 dates have one event and 39 dates have 4 or more, which is why same-date entries are capped at 3 (L267). Worst single trades were -37.6%, -34.2% and -30.1%; at 6% per leg the worst trade costs about 2.3% of equity. A sequential date-level simulation at 6% per leg and 3 per date, after 0.35% round-trip cost, gives max DD about 16% and about 660 trades. Independent evidence is 493 event dates, not the per-leg count (L238). As a negative control the same rule was measured on Swiss large caps (CHEQ: +0.34% raw, +0.13% excess, and negative at 3 sessions), which is why Switzerland was not chosen. COST ARITHMETIC: KREQ charges 0.0186% on buys and 0.2186% on sells (the sell side includes the securities transaction tax), plus about 0.05% half-spread per side on these liquid large caps, so a round trip costs about 0.34%. Expected gross capture is the median +0.80% / mean +1.55%, so net is about +0.45% to +1.2% per trade, at least 2x cost on the mean. The peer excess alone (+0.70%) also clears 2x cost. EXECUTION: KREQ is a session venue. The signal is computed on the close of day t, and a market order decided then fills at the next session's open (next_open_fill, supported; no same-close fill is required). The exit is decided at the close of the 5th held session and fills at the following open. Daily price limits are enforced by the platform: a market order into a locked limit is refused, and that is accepted as part of the test. The venue cannot short, so the design is long/flat and unhedged. Risk control is route (b): a time exit plus capped sizing, 6% of equity per leg and at most 36% gross. No leverage. Metrics are in KRW.

Implementation

33-name KRX large-cap common-share panel on KREQ 1-DAY bars (primary 005930, 32 extra legs, every name both a tradable leg and a peer). At each session close, every listed name with bars on t and t-1 gets r = ln(C_t/C_{t-1}) and a peer return P = equal-weight mean of all OTHER names with a bar that day. beta = cov(r,P)/var(P) and sigma = std of (r - beta*P) come from the name's last resid_lookback (r,P) sessions ending at t-1, and z = (r - beta*P)/sigma. A name with z <= -z_threshold that is not held and has no entry in flight triggers. At most max_new_per_day names enter per date, most negative z first, and open plus in-flight legs stay <= max_concurrent. The market BUY is sized at position_pct x mark-to-market equity / close in whole shares, and fills at the next Seoul open (next_open_fill). The exit is a fixed clock: at the close of the hold_sessions-th session of that name's own bars (the fill session counts as 1), a market SELL of the full leg fills at the next open. No stop, no target, no signal exit. Long/flat, unlevered CASH account, KRW metrics.

Implements the hypothesis as written, using the promoted ASX/TSX peer-residual panel structure. Eligibility is point in time: a name needs resid_lookback completed (r,P) sessions, i.e. resid_lookback+1 sessions of bars, and must be listed, so LGES, SK Square and HD Hyundai Heavy enter only once they have that history; point_in_time_universe=True stops later listings from clamping the window. One lookback drives both beta and sigma. Sigma is the std of the residuals r - beta*P over the window, where beta is the cov/var slope estimated on that same window ending at t-1. Positions exiting at today's close still count toward max_concurrent until they fill, as in the promoted siblings, which is slightly conservative. No limit-down filter: the platform refuses orders into a locked limit. Risk route (b): time exit with 6% of equity per leg and at most 6 legs (36% gross), unlevered. Worst measured trade is about -38%, roughly 2.3% of equity, inside the profile's 15% per-trade cap. Rolling windows are rebuilt from the bar buffers after a restart or a skipped session, so a restart changes no decision. Clamps match the optimization-plan domains and are declared in _param_bounds.

Verification Results

None required.

min_bars_required() is overridden to resid_lookback+2 (static base_shadow warning). The override is deliberate and allowed: the base docstring says 'Override as needed', and the 500-bar buffer covers the maximum of 122. Bars below the minimum skip calculate_signal, but the first processed bar triggers _rebuild() from the buffers, so the windows come out complete.

Optional: exclude legs with _close_pending from held if exact slot accounting is wanted.

held counts legs whose exit SELL was submitted at today's close (they are still open until the next open). On an exit day this can reduce the free slots by up to the number of exiting legs, which is slightly more conservative than 'open-position count <= max_concurrent' after fills. The developer disclosed this, and it matches the promoted sibling panels.

Survivorship: the panel is today's KOSPI large caps, so judge the result on the excess over the peer basket, not on the raw return. The pre-registered excess (+0.70% per 5 sessions) is only about 2x the ~0.34% KREQ round trip, and 2022 was negative. Market-wide crashes cluster entries (3 per date cap). Separately, the sandbox took 221 s for 1,477 bars (runtime_review_required). Per-bar work is O(33 x lookback), so the cost is probably the 33-leg engine load rather than the strategy, but it is worth a look before 30-150 trials per window.

Backtest Review

Sharpe: 0.59
Total return: 26.66%
Max drawdown: 12.86%
Trades: 637
Win rate: 49.6%
Profit factor: 1.34

The trades implement the hypothesis: every leg is long, no date has more than 3 entries (the cap binds), 392 of 504 legs last 7 calendar days (5 sessions; longer ones span holidays), and size is 4-6% of equity per leg. Implied deployment is 14.71 / (504 x 0.504) = about 5.8% per trade, which matches the 6% rule, so the 12.9% drawdown is measured at the intended size.

Economics are positive after KREQ costs: avg_trade_return_pct 0.504% against the 0.45% KREQ floor, PF 1.24, Sharpe 0.48, total return +14.7%, max drawdown 12.9%, inside the pre-registered 22% constraint.

Worst leg was -30.4%, about 1.8% of equity at 6%. This matches the risk-route (b) design (time exit plus capped sizing).

The edge is thin and depends on the regime. Mean net trade return by entry year: 2018 +0.81%, 2019 +0.44%, 2020 +0.01%, 2021 -0.11%, 2022 -0.81%, 2023 +1.34%, 2024 +1.76%. Most of the profit comes from 2023-24, which are the years the survivorship caveat flags (today's winners in the panel). Benchmark-relative alpha is withheld in this view, so the peer excess the hypothesis says to judge on cannot be confirmed here. The analyst must read it after optimization.

The realized 0.50% net per leg is below the pre-study's implied about +1.2% net (1.55% gross mean minus 0.34% cost). Per-trade return is only about 1.1x the venue floor, not the 2x economics target. That soft floor will need to be met or waived later.

The March 2020 crash cost -6.5% in one month and 2022 cost -5.7%. Market-wide shocks still cluster losses despite the cap. The 949-day drawdown duration is long.

QA's survivorship concern is partly supported (raw returns rely on 2023-24); QA's runtime concern holds: 405 s per backtest x 90 trials per window is heavy but feasible.

Backtest and paper results are hypothetical. Trading involves risk of loss.