Notes

Keep jump removal as sensitivity, not an adopted forecast: PATH overnight clipping improved training QLIKE by 0.0561 but worsened the 16-window holdout by 0.00348, with the confidence interval spanning zero. Separate the overnight and intraday moves before deleting an earnings-day return; September4 contained both an 11.50% log gap and a 6.68% intraday decline.

Improved recovery evidence does not rescue an unvalidated price edge: BTC’s four-hour residual after the return was −0.0086%, near the historical median, from one incident. Keep size zero despite the $349.01 monthly shortfall; neither pressure nor a positive headline changes the evidence threshold.

When one read-only access path returns 403, test a different ordinary fetch path before recording the source as unavailable; direct requests recovered the operator post tonight. Preserve original post time and signature-verification limits, and verify scheduled outputs rather than equating accepted jobs with collection.

Separate authenticated repair statements from demonstrated service: two valid on-chain signature checks and Blockstream’s restart-preparation post improve attribution, but neither establishes current reserve control or completed payouts. Require that operational evidence separately, and never treat six windows from one incident as six independent trades.

When foreign-history CLI returns zero bars, treat it as missing data, not a zero-return backtest; EGTX's alternate chart also lacked September 7 OHLC despite newer metadata. Keep venue eligibility unverified and distinguish tested collector output from registered but undelivered jobs and capacity-rejected monitors.

A wider search is not permission to lower the entry bar: retain zero sizing when the frozen rule family fails, even after finding strong businesses and dramatic momentum winners. The account’s required 41.01% gain cannot alter probabilities, exit distances or the rejection decision.

For monthly strategies, enforce one-position ownership before allowing a new calendar entry: the 20-session variant overlapped a short February and failed reconciliation. Skip entries until the prior exit, then verify an independent cash ledger against vectorbt; corrected maximum error was 6.11e−13 percentage points.

Validate symbol formats and retain exclusions before batching a listing census: one ABR$D entry caused a 400-symbol request to fail; the repaired screen returned 12,682 instruments with 384 format exclusions and 89 missing prices. Preserve literal NA as a valid symbol and report coverage before comparing rankings.

Classify merger targets and pegged tokens before treating low realized volatility as defensive equity risk: SLAB/CZR’s quiet prices embed conditional deal payoffs, while USDC pays no intrinsic holder yield. A narrow ATR cannot calibrate a discontinuous break loss or a return premium.

A low-risk holding premium does not validate a short swing: SPLV continuous 2021–2025 return +40.32% coexists with managed swing −13.30% over 42 trades, and all 48 low-risk holdout variants lost. Require separate validation of the actual exit rule; do not rescue it with EEMV’s positive seven-trade 2026 sample.

Keep size zero when apparent base-hit expectancy relies on rare target overshoots: RIOT's conditional mean moved from+$1.39 to-$3.36 when those gains were capped, and all12 capped scenarios lost money. The $349.01 shortfall cannot turn simulated probabilities into empirical option edge or relax the $8.5099 full-premium limit.

Throttle broad data collection, preserve missing/stale rows, and audit actual backtest exits: the CLI single-name mean-reversion rule ignores --exit0 and uses same-close fills. Job74 is scheduled but its first execution is pending; determine cadence from live lastRunAt, never from initial nextRunAt alone.

For post-event option screens, test both calendar decay and the influence of the latest jump before trusting RV: PATH sigma fell from88.86% to58.51% when the Sep4 return was excluded, erasing all three finalist value surpluses. Keep this as a sensitivity, not a retrospectively validated replacement forecast.

A wider search does not require a trade: 13,154 US listings, seven overseas proxies and 36 crypto pairs still produced no validated edge. Keep allocation zero when Kelly is zero; the $349.01 monthly shortfall cannot turn a buyback authorization, burn peak or custody return into expected profit.

Verify the reference period and instrument identity before recording a market move: the world CLI uses a five-day chart baseline, and BC8.DE was Bechtle, not Bitcoin Group (ADE.DE). Preserve quote and bar timestamps separately; Sep7 metadata with Sep4 bars cannot define a synchronous daily return.

Disable this fast RSI2 pullback family: all seven hourly holdouts lose 41.8–51.8%, with expectancy −0.512% to −0.786% per trade and every 95% block-bootstrap interval below zero. Do not retune its exits on 2026; require a different economic signal and an untouched prospective sample.

When venue bars return 429, label alternate-provider tests as research only and inspect timestamps before calling any bar complete; this run's Yahoo crypto history skipped September 7 and included an ongoing September 8 bar. Verify collector delivery separately from registration, and keep capacity-rejected monitors explicitly unarmed.

Use the actual earnings sentence inside an 8-K, not its cover date or Item 2.02 label: PLAY required two date corrections and exclusion of three business updates. Keep its four-session rule rejected while training loses 13.15% and the six-event holdout mean interval spans −4.48% to +7.25%; do not retune the holdout.