Connors RSI(2) Pullback in the Index-ETF Domain It Was Designed For: Long-Only SPY/QQQ/IWM/DIA When Close > SMA200 and RSI(2) < 5, Exit on Close > SMA5 or 10-Session Clock, Published Parameters Frozen as a Post-2009 Out-of-Sample Test (USEQ 1-DAY)

Hypothesis

Implements Connors' published rule with the article's parameters frozen (SMA200, RSI(2) <= 5, exit on close > SMA5), plus the hypothesis's 10-session clock, on the four broad index ETFs. Each ETF is its own sleeve; an extra leg acts only on a bar stamped at the primary session close. use_trend_filter=false gives the pre-registered secondary variant without SMA200. Iteration 3: v2 was profiled in the actual sandbox runner. Run time is dominated by NautilusTrader's catalog.bars() fsspec glob over about 126k bar directories on an overloaded host (load ~34). Per-bar strategy work is O(1) per leg, and the strategy ran to completion (1406 bars, 28 SPY entries). The trading logic is unchanged; only documentation changed. The timeout is infrastructure (catalog scan / host load) and should be retried or fixed on the platform side, not in strategy code.

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 SPY.USEQ-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

Connors RSI(2) pullback, long-only, on four independent USEQ daily sleeves (SPY primary; QQQ/IWM/DIA extras). Entry: close > SMA200 and Wilder RSI(2) <= 5. Exit: close > SMA5 or after 10 sessions. Each sleeve is 25% of equity (max gross 1.0x, inside the 2x Reg-T cap). Whole-share sizing.

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 SPY.USEQ-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).

Outcome Summary

When a sandbox timeout comes from the catalog/data-loading layer and not from per-bar strategy cost, changing the strategy code will not fix it; it needs an infrastructure fix or a manual escalation, or the verification loop keeps abandoning a valid hypothesis before any backtest runs.

The strategy was abandoned in a verification loop after 3 iterations. Each attempt failed Layer 3 (sandbox backtest) because the smoke test ran past the 300s wall-clock limit on a ~20000-bar SPY.USEQ 1-DAY window. The developer's own profiling blamed ParquetDataCatalog bar-file globbing (~126k bar dirs on a heavily loaded host), not per-bar strategy work.

Test Connors' published RSI(2) pullback rule, with its parameters frozen, as a long-only independent sleeve on each of SPY, QQQ, IWM and DIA (USEQ, 1-day bars): buy when close > SMA200 and Wilder RSI(2) <= 5, and exit when close > SMA5 or after 10 sessions. The rule relies on short-horizon negative autocorrelation in index returns, and data from 2009 onward serves as a pre-registered out-of-sample test.

No performance numbers exist. The backtest report is empty and optimization, backtest review and analyst review never ran. The only figures come from a note in the iteration-3 code: the sandbox smoke test ran 1406 bars and made 28 SPY entries before it hit the time limit.