AhDualListingDislocationHarvest

A-H Dual-Listing Dislocation Harvest: Buy the Cheaper Listing of the SAME Company Across HKEQ and CNEQ (10 verified A/H twin pairs, daily bars, z-score of log(A/H) vs its own 120-session norm, single-leg long, 3-parameter)

A CROSS-VENUE, CROSS-MARKET RELATIVE-VALUE strategy on 10 Chinese companies that are simultaneously listed as an A-share on the mainland (CNEQ, closes 07:00 UTC) and an H-share in Hong Kong (HKEQ, closes 08:00 UTC). The two lines are claims on IDENTICAL corporate cash flows, but capital controls and Stock-Connect quotas keep the two investor bases segmented, so the price ratio drifts far from its own norm and then re-converges. Because BOTH venues are CASH accounts (no shorting, no leverage), the strategy does NOT run a two-leg spread: it holds exactly ONE leg — the listing that is currently cheap relative to its twin — and uses the other leg purely as a SIGNAL SOURCE. One fee-paying leg, one round trip, no borrow. Universe (primary first, then extras; each HK ticker is paired with its mainland twin): 2318.HKEQ/601318.CNEQ (Ping An), 1398/601398 (ICBC), 2628/601628 (China Life), 6030/600030 (CITIC Securities), 3968/600036 (China Merchants Bank), 0857/601857 (PetroChina), 0386/600028 (Sinopec), 1088/601088 (China Shenhua), 2601/601601 (China Pacific Insurance), 0998/601998 (CITIC Bank). All 20 instruments were VERIFIED present in the catalog with 1-DAY bars from 2018-01-02 to 2026 (~2,100 sessions each, ~8.7 years). This is the first hypothesis ever proposed on HKEQ (0 of 3,068 experiments) and the second on CNEQ (1 of 3,068) — the two emptiest usable venues in the entire data estate, both with working session calendars (XHKG, XSHG) and real statutory cost models in venue_configs.py.

Iteration 2 addresses the single issue named in the feedback: the optimizer selected a non-trading corner of the search space (entry_z pinned at 3.9964 on its 4.0 upper bound, giving 21 entries in 8.7 years and zero inside any 377-day validation window, hence OOS Sharpe literally 0.0 and a 3-trade holdout). The signal logic, pairs universe, position_frac, alignment and every other parameter are untouched — the base config already produced 507 trades, 15/20 legs profitable, avg_trade_return_pct 1.258% (~4x the HKEQ round trip), DD 10.3%, 8/9 positive years. The only edits are the three _param_bounds narrowings requested, mirrored exactly in the code's clamps so the optimizer never tests a value the code silently rewrites: entry_z [0.5,4.0] -> [1.0,2.5], lookback_days [40,300] -> [120,260] (removes the only sensitivity cliff: 0.245 at 96 / 0.271 at 108 vs 0.644-0.721 at 120-144), exit_z [0.0,1.2] -> [0.15,0.75] (drops the degenerate exit_z=0.0). No other change.

A/H dual-listing dislocation harvest across 10 verified twin pairs (H-share on HKEQ, A-share on CNEQ). For each pair it tracks r = log(H/A) against that pair's own trailing 120-session norm; when r sits entry_z sigma below the norm the H line is the cheap claim on identical cash flows and the H leg is bought, entry_z above and the A leg is bought. Exactly ONE long leg per pair (both venues are CASH, long-only, no leverage) — the twin is a pure signal source, so there is one fee-paying round trip per dislocation. Exits when the ratio reconverges inside exit_z, on a 90-session time stop, or on a 20% leg stop; a 1-session minimum hold respects CNEQ T+1. Point-in-time discipline: the ratio series is stamped at each 08:00 UTC HK close using the 07:00 UTC CN close already known at that instant.

The prior iterate feedback was fixed at the right layer: _param_bounds now caps entry_z at 2.5 (was 4.0) and floors lookback_days at 120, so the optimizer can no longer pin entry_z on a non-trading bound. At the default entry_z=1.5 the universe produces 507 closed trades over 2022 sessions — every IS/OOS/holdout window will now contain trades.

Large, well-spread sample: 507 trades across all 20 legs (288 CNEQ / 219 HKEQ), metrics_reliable=true, median hold 20 days, 8.7 years of daily bars; 15 profitable names vs 5 losers, largest leg only 18% of gross.

Trades match the hypothesis: 507/507 LONG (both venues are CASH), one leg per pair, zero same-day round trips (min hold 1.0 day) so CNEQ T+1 is respected.

Per-trade edge far above the HKEQ floor: 1.25% vs 0.30%; PF 1.47, win rate 62%, DD 10.3%, impact only 7.3% of gross, capacity $18.7M, profitable in 8 of 9 calendar years with no single-year concentration.

VERIFIED engine gap: commissions were not charged. total_commission=0.0 and all 507 trade records show commission 0.0, while venue_configs sets HKEQ 0.11%/side and CNEQ 0.05%/side. Cause located in code — BacktestRunner._apply_fees() returns early for any NT Equity instrument (written for commission-free USEQ), and HKEQ/CNEQ instruments are Equity too. Estimated uncharged cost on actual notionals ≈ $10.6k vs $80.4k gross PnL (~13%, ~0.17%/trade). Net per-trade edge ≈ 1.08% — still ~3.6× the floor, so not disqualifying, but every headline number is optimistic by that amount.

Sharpe 0.883 with a wide CI (0.24–1.55) and PSR 0.86 — deflation over ~225 trials will be a real hurdle at ANALYZING.

Sortino 0.47 « Sharpe 0.88, skew −0.41, kurtosis 7.0, 13 consecutive losers, 617-day max DD duration — the usual convergence-trade shape (gaps widen before they close).

alpha/beta/IR are null, and 4.2% CAGR is modest; the optimization must show dislocation capture rather than long-China-financials beta.

The 3-phase run did not test this premise out-of-sample — it tested a non-trading corner of the search space. entry_z was pinned to 3.9964, ON its upper bound of 4.0; at |z|~4 over a 123-session norm the 10-pair universe fires 21 times in 8.7 years and ZERO times inside each 377-day validation window, which is exactly why all three OOS Sharpes are literally 0.0 and holdout has 3 trades. Meanwhile the base config is broad and healthy (507 trades, 15/20 legs profitable, max leg share 18.2%, both venues positive, 8/9 years positive, avg_trade_return_pct 1.258% ≈ 4x the HKEQ round trip, DD 10.3% vs a 22% cap, Sharpe CI lower bound > 0). The fix is to the SEARCH SPACE, not the logic. Make exactly three _param_bounds edits: (1) entry_z [0.5,4.0] -> [1.0,2.5]; (2) lookback_days [40,300] -> [120,260], removing the only cliff (0.245 at 96 / 0.271 at 108 vs 0.644–0.721 at 120–144); (3) exit_z [0.0,1.2] -> [0.15,0.75], dropping the degenerate exit_z=0.0. Change nothing else — not the signal logic, the pairs universe, position_frac, or the point-in-time CN-close-before-HK-close alignment. This is attempt 2 of 2 and it is the last: it must produce a selection that TRADES IN EVERY VALIDATION WINDOW. If it does and OOS Sharpe is still ≤ 0.5, or PBO stays > 0.5 with trades present, that is a genuine out-of-sample measurement and the correct outcome is abandon, not another bounds tweak. Full write-up in workspace/discussions/48542967-d379-478d-bc74-478614bfbc04/iteration_1_feedback.md.

Benjamini-Hochberg at q=0.10 over 288 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.