OKX Cross-Sectional 52-Week-High Anchoring Spread: Dollar-Neutral Long Names Nearest Their Trailing High / Short Names Furthest Below It (15 OKX USDT Swaps, Daily Bars, Weekly Rebalance)

Hypothesis

Fills the long_short (16.5%), OKX (1.9%) and multi-instrument buckets and avoids BINANCE/BTC concentration. The 52-week-high anchoring family has no prior hypothesis. It avoids L111 (no supplementary gate), L179 (rank-based trigger; the weekly clock only sets cadence) and L102 (no options). Legs were verified in the catalog per L186. Leverage is deliberately absent from optimization_plan to avoid the Layer-1.5 'fixes leverage' mismatch; gross exposure is set by per_leg_weight x K.

verification_loop: sandbox_timeout: Verification failed (Layer 3 — sandbox backtest): smoke test exceeded the 300s wall-clock limit on a ~20000-bar window (2400 days of BTCUSDT.OKX-1-DAY-LAST-EXTERNAL). This almost always means per-bar work that scales with history — e.g. rescanning the full funding/supplementary series, or rebuilding a list and calling min()/sorted() inside calculate_signal()/on_bar() on every bar. Precompute sorted timestamp arrays ONCE in __init__ and use bisect, or cache lookups keyed by timestamp, so per-bar cost is O(log n) not O(n).

Implementation

Weekly Monday UTC close ranking of 15 OKX swaps by close divided by trailing daily high. Holds equal-dollar long top-K and short bottom-K targets, resizing retained legs in place, with daily-close leg and book stops.

Preserves the previous strategy's signal, universe, schedule, risk thresholds and research parameter permissions. Replaces repeated max(buffer) scans with mathematically equivalent bounded monotonic queues: amortized O(1) per incoming bar and O(1) per PTH lookup; sorting remains weekly over 15 names. No performance test was run, so the 300-second timeout is not confirmed resolved. The previous implementation was already bounded and used active-order queries; another timeout should trigger engineering profiling of data loading, funding accrual, execution and reporting rather than another speculative signal edit. No final-exam data was read. The requested outbox file could not be written because this session permits only filesystem reads; the complete task-bound JSON is returned here.

Verification Results

sandbox_timeout: Verification failed (Layer 3 — sandbox backtest): smoke test exceeded the 300s wall-clock limit on a ~20000-bar window (2400 days of BTCUSDT.OKX-1-DAY-LAST-EXTERNAL). This almost always means per-bar work that scales with history — e.g. rescanning the full funding/supplementary series, or rebuilding a list and calling min()/sorted() inside calculate_signal()/on_bar() on every bar. Precompute sorted timestamp arrays ONCE in __init__ and use bisect, or cache lookups keyed by timestamp, so per-bar cost is O(log n) not O(n).

If the book stop is meant as a cooldown, skip that day's rebalance after a stop has fired, or require N weeks flat.

The book stop calls _flatten_all(), but on a Monday close _rebalance runs in the same calculate_signal call. It sets new targets right away, and those are worked on the following bars. So the 12% book stop can be followed by immediate re-entry into a near-identical book, which defeats much of its purpose. The hypothesis does not specify stops, so this is not a mismatch.

No change required. Optionally attach reduce-only stop_market orders to short legs, where squeezes are the main tail risk.

Leg and book stops are checked only on daily closes, with no intrabar protection. A 25% leg stop on an 8% leg caps the planned loss at about 2% of equity per leg, but gaps can exceed it. This is documented and acceptable on 1D bars.

Acceptable for the current universe. Consider ranking over the names that are available (needing at least 2K) instead of requiring all 15.

Every rebalance flattens the whole book if any one of the 15 names lacks a full lookback or has a bar older than stale_bar_days. A single delisted or halted OKX swap would therefore stop the strategy permanently. The base alignment barrier also skips any day on which one leg is missing a bar.

Profile _work_pending/_risk_checks. Return early from _risk_checks when no positions are open, and call _close_pending only for names that have pending items.

The sandbox needed 261s for 1480 days of 15 daily legs, and runtime_review_required=true. The strategy's own per-bar work is bounded: fixed deques, max() over at most 365 highs only when ranking, and a 15-name loop. The cost therefore probably comes from repeated cache.orders scans in _close_pending/_entry_order_inflight and engine overhead, not from history rescans. The full span is 1849 days, so a single backtest should finish, but optimization trials will be expensive.

Smoke test: Sharpe 0.33 with CI [-0.64, 1.22], PF 1.09, max DD 31.9%. The stressed-vol tercile lost 12.7%, and rate_shock_2022 lost 19.2%. Funding is NOT folded (funding_events_available=0), yet OKX funding exists from 2022-01. A long-high/short-low book held about 38 days per leg tends to sit long the crowded names that pay funding, so the full backtest must account for funding. The ledger is also inconsistent: avg_trade_pnl_usd is +$62 while avg_trade_return_pct is -0.29%. In-place resizes and netting flips probably distort the per-position % basis. Reconcile this before using the % figure against the 0.15% floor. Runtime was 261s for 1480 days, so profile the per-bar overhead before optimization.

Analysis

Iteration History

Verification failed (Layer 3 — backtest process): the strategy could not be loaded/built in the backtest process: in _run_backtest supp = _resolve_supp(input_config) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/strategy-factory/src/backtesting/subprocess_entry.py", line 118, in _resolve_supp supp = load_for_subprocess( ^^^^^^^^^^^^^^^^^^^^ File "/opt/strategy-factory/src/backtesting/supp_collector.py", line 664, in load_for_subprocess supp = asyncio.run( ^^^^^^^^^^^^ File "/usr/lib/python3.12/asyncio/runners.py", line 194, in run return runner.run(main) ^^^^^^^^^^^^^^^^ File "/usr/lib/python3.12/asyncio/runners.py", line 118, in run return self._loop.run_until_complete(task) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete File "/opt/strategy-factory/src/backtesting/supp_collector.py", line 683, in _ensure_then_load_supp await _ensure_perp_fu

Outcome Summary

A 15-instrument strategy has to keep its per-bar work small (bounded rolling maxima, ranking only on the rebalance bar, and no per-bar cache/order scans across the universe), and that cost should be checked against the 300s sandbox budget before iterating. In this case even the hypothesis's explicit warning and the monotonic-queue design did not get it under the limit.

After 6 iterations it was abandoned in the verification loop. The Layer 3 sandbox backtest went over the 300s wall-clock limit on a ~20,000-bar window (2400 days of BTCUSDT.OKX 1-DAY bars), so the backtest-review, analyst and risk stages never ran.

A dollar-neutral weekly long-short spread across 15 OKX USDT swaps on daily bars. It went long the names closest to their trailing 52-week high and short the names furthest below it (George & Hwang anchoring ratio, close / trailing max high), as opposed to ranking by return momentum.

No performance figures exist. The strategy never passed verification, so there is no backtest report and no optimization report: no return, Sharpe or trade count.