EXP-015 — Cross-sectional short-term reversal

EXP-015 — Cross-sectional short-term reversal

• Status: pre-audit complete · data_hold (2026-09-17)
• Contract: common contract
• 1st hypothesis: among liquid stocks in an uptrend, buying the relatively weakest over the last 5 days earns more after costs than random selection.

Universe

• Common stocks only; ETF/ETN/preferred/SPAC excluded. At least 252 trading days of history, median trading value over the previous 60 trading days observed up to that day ≥ 1,000,000,000 KRW, positive volume on ≥55 of 60 days.
• At each month-end, the 100 qualifying stocks with the highest median trading value over the last 60 days are selected and fixed for the next month. Ties by symbol code; if eligible names are short, only the actual count is used, and new entries halt in months with fewer than 50.
• Original actual trading-value data are used first. It is not assumed that adjusted price × raw volume restores the original trading value.
• Without a point-in-time master, the work is limited to exploration within the current snapshot's common stocks. Stocks are not filtered by requiring data to exist into the future, which would add survivorship bias.
• The top 100 is not a ranking over the whole future period, so the calibration period also uses only data up to that point.

Signal and operation

• Each day t, sort non-held stocks with close > SMA200 by r5 = close_t / close_(t-5) - 1 in ascending order.
• No absolute thresholds such as RSI<5 or r5<0. In a broad rally, stocks with positive returns can also be selected. Their frequency and contribution are reported so it is not misread as 'buying only falling stocks'.
• K=5, a 19% budget per stock, filling only empty slots at the next day's open. Holdings are not replaced just because their rank changes.
• Enter at O1, exit at O6 5 trading days later. The scheduled exit date is set by the market calendar and waits during trading halts. No same-day re-entry into the same stock. This is long-only short-term reversal and is not called a long-short factor return.
• Ties in new-candidate rank go by symbol code. Entry rank is not re-decided by same-day return/volume.

Comparisons and separating the changed factor

• 1st comparison: an account that fills empty slots uniformly at random instead of by rank, in the same trend and universe each day. Each random account keeps its own holdings and cash path. It is not adjusted to the strategy's fill count or ex-post performance.
• The mean daily log-return of 200 accounts with fixed seeds (15000~15199) is used as the 1st control. The 5/50/95th percentiles of the distribution and the strategy's actual position are also published. Unfavourable accounts among these seeds are not chosen as the comparison.
• An account running the absolute RSI<5/RSI70 exit in the same monthly universe is a reference. It is not directly subtracted from EXP-013's performance on a different universe and called a rank effect.
• Cash, market95 and a monthly equal-weight account on the same universe are shown separately as well.

Preregistered matrix

• Rank {r5, RSI2} × K {5,10} × 3 periods × 4 costs = 48 strategy cells.
• 1st cell fixed at r5/K5/history_B/proxy_slip2. RSI2 ranking is a definition-robustness diagnostic; picking the best value is forbidden.
• Random controls for K5 only: 200 seeds × 3 periods × 4 costs = 2,400 paths. Indicator calculations and price paths are reused.
• 12 reference cells for the absolute RSI rule, 36 cells for cash/market95/equal weight. These numbers are a future cost budget; nothing is run this time.
• Superiority of 10 slots alone does not support a claim of a 'reversal selection effect'. It is an exposure/diversification diagnostic.

Sample and verdict

• In history_B: completed trades ≥200, union of entry days ≥100, trading months ≥24, selected stocks ≥30. If the calendar lacks the required data, the bar is not lowered; it is hold.
• Research promotion: actual account net return >0, lower CI bound of the 1st log-return difference vs the random mean >0, positive own net return and excess point estimate under both cost stresses, positive history_A excess point estimate.
• Positive after excluding the top stock/year P&L, and losses from unfilled/rejected orders and trading halts not omitted.
• If higher volume only raises gross profit while net return disappears, the after-cost operating contract fails.
• Even if it underperforms the RSI reference over the same period, record it separately and measure the correlation between strategies. Do not optimize weights after the fact to build a blended portfolio.

Pre-implementation validation and planned outputs

Synthetic tests cover month-end → next-month universe application, liquidity and new-listing warm-up, rank ties, no re-buying held stocks, separation of close/open timing, keeping slots for unfilled exits, seed reproducibility and independent cash per random account. Later outputs: rank/universe/candidate/rejection ledgers, trades/NAV, random control distributions, turnover and slippage breakevens, monthly/yearly and per-stock contributions and a verdict table. Performance has not been measured yet.