Replicate RecessionAlert indicator: WLEI second-derivative warning

Objective

Independently replicate the distinct RecessionAlert row 2nd derivative of WLEI posts warning from the checked-in canonical snapshot, without a live-site request, credentials, or private account data. Because the source omits an executable series and action rule, this result follows the repository's best-effort research-campaign policy rather than stopping at a gap-only outcome.

Source and disclosed rule

The canonical input is the manifest-selected page body in read-only data/recessionalert.sqlite, identified by research/artifacts/recessionalert/manifest.json and research/findings/recessionalert_inventory.md. The redacted source location is https://recessionalert.com/2nd-derivative-of-wli-posts-warning/; the selected page SHA256 is sha256:9a47b636c5273dbe77dc525e0fdaafc9e1e0d8927ea71d86b3f7b23fa3a663a2, fetched at 2026-08-22T23:46:05Z, with content length 80,946 bytes. The article date is 2018-11-16T10:35:28-0500. The crawl log retains an earlier same-URL body (sha256:b22560554fc2fc770e999465353245ca9efdb796a201a83a6f55bb138785081c); the later body is the manifest-selected revision.

The article describes WLEI as essentially a first-derivative/rate-of-change indicator. It says a second derivative can be formed from the historical percentage of WLEI readings above the current reading, interprets that context as the percentage of WLEI components in recession territory (WLEI Diffusion), and discusses divergence followed by Diffusion catch-up. It presents ECRI WLI as a possible first warning and WLEI as a second warning, advising attention and consultation of more robust models when both flag recession.

This is a warning concept, not a disclosed trade rule. The source gives no buy-only, sell-only, dual-sided, risk-on, risk-off, target-weight, universe, rebalance, exit, or re-entry semantics.

Best-effort proxy reconstruction

The canonical row has no source-linked WLEI/WLEI Diffusion/ECRI WLI observations or chart payload. The longest defensible checked-in public substitute is monthly FRED INDPRO (1919-01 through 2026-06, 1,290 observations). The inferred fixture research/signal_spec/fixtures/recessionalert-wlei-second-difference.json computes a rolling historical rank and its second difference, with signal_value < threshold as an inferred warning condition. It is explicitly not an official WLEI or RecessionAlert parity implementation.

Actual cached Yahoo OHLCV prices provide SPY as the risk-on benchmark (1993-01-29 onward) and IEF as the risk-off sleeve (2002-07-30 onward). Macro period-start dates are normalized to month-end; warm-up and missing states remain unavailable rather than risk-on. Primary results apply at least one available-bar lag. Zero lag is reported only as timing sensitivity; same-close execution is unsupported.

The exploration also loaded UNRATE (1948-01 onward) and a transparent GS10 - TB3MS curve proxy (common span 1953-04 onward) for diagnostic combinations. No live FRED/Yahoo refresh, RecessionAlert workbook substitution, synthetic price backfill, constituent rebuild, or third-party acquisition was used.

Tested variants and common-window results

All four rows are inferred, exploratory variants evaluated on the common SPY/IEF window 2002-08-31 through 2026-06-30 (287 months):

All lagged proxy overlays underperformed buy-and-hold SPY cumulatively. The three lag-1 variants had shallower descriptive maximum drawdowns than SPY; the stricter lag-2/threshold/holding neighbor did not. These are exploratory metrics, not a promotion gate or claim about native RecessionAlert performance.

Diagnostic results are persisted in research/experiments/1335-state-space.json and research/experiments/1335-diagnostics.json. On the proxy labels, eight INDPRO YoY contraction onsets had a prior six-month target warning, but active-warning months had a higher mean forward six-month SPY return (0.064304) than clear months (0.052260). Curve-inversion overlap was 16 months for target-plus-curve warnings; target-plus-UNRATE overlap was 36 months. Lag sensitivity was material: lag-0 diagnostic overlay cumulative return 8.862957, lag-1 primary causal 5.355754, and lag-2 causal sensitivity 2.265437.

Causal assumptions and native gaps

The native source span is one dated HTML article snapshot, not an observation history. Underlying WLEI/WLEI Diffusion/ECRI WLI publication lag, release clock, timezone, point-in-time availability, vintage, and revision policy are unknown. The 2,836-day article-to-archive interval is snapshot delay, not economic publication lag.

The inferred proxy uses actual stored period-end values, one or more available-bar lags, and no same-close execution. Missing and warm-up observations stay unavailable; they are not imputed, forward-filled, false, neutral, or risk-on. FRED values can be revised, and monthly INDPRO is not weekly WLEI.

Explicit remaining gaps are:

• source-linked weekly WLEI, WLEI Diffusion, and ECRI WLI observations and native observation span;
• the exact historical rank window, denominator, date/tie convention, score range, smoothing, and missing-value behavior;
• numeric warning threshold, divergence tolerance, persistence, confirmation, and state transitions;
• weekly release calendar, publication clock, timezone, information cutoff, vintage, and revision policy;
• buy/sell/hold meaning, target weights, investable universe, rebalance cadence, holding period, exits, re-entry, costs, and slippage;
• the three referenced chart payloads and proof that any checked-in workbook belongs to this distinct row;
• independent recession and market outcome windows sufficient for confirmatory validation.

Files and reproducibility

• research/findings/specs/recessionalert-wlei-2nd-derivative.json — source rule, provenance, native span, lag boundary, and gaps.
• research/signal_spec/fixtures/recessionalert-wlei-second-difference.json — inferred executable SignalSpec.
• research/research_campaign/recessionalert.py — read-only proxy loading and variant evaluation.
• research/experiments/1335-state-space.json — four-variant trial ledger.
• research/experiments/1335-diagnostics.json — regime, event, combination, and lag diagnostics.
• research/experiments/1335-synthesis.json — ranked candidate frontier and next questions.
• research/campaigns/recessionalert-wlei-2nd-derivative/campaign.json and lab-series.json — canonical research-library inputs.
• research/experiments/1335.ipynb — executed offline notebook with tables and signal chart; rendered to docs/experiments/1335-analysis.html.

The public research page is generated at docs/research/recessionalert-wlei-2nd-derivative.html with payload docs/site-data/research/recessionalert-wlei-2nd-derivative.json. Public outputs contain summaries, derived metrics, hashes, provenance paths, causal assumptions, and gap notes only; cached HTML/chart/workbook bytes, credentials, cookies, and private account data are excluded. The proxy candidates remain Other / Research evidence and are not registered as an official Other / Indicators signal.