Pipeline Retrospective — 2026-04-20 #

Cycle: 2026-04-20 (Q2 bi-monthly, first cycle after 2026-04-03)

Prediction Resolution Status #

Zero predictions resolvable this cycle. The prior cycle's 7 predictions (2026-04-03) are only 17 days old — minimum resolution horizon is 30 days (earliest resolvable 2026-05-03).

IC Computation & Weight Adjustments #

Not computed this cycle. Minimum threshold is 10+ resolved predictions. Currently 0 resolved, 14 pending.

Weights remain at defaults (v1):

• graham_pct: 0.30
• magic_formula_pct: 0.20
• llm_pct: 0.30
• schilit_clean_pct: 0.10
• reddit_divergence: 0.10

This Cycle's Additions #

7 new predictions recorded for 2026-05-20 resolution:

• PRU.L (BUY, conviction MEDIUM-HIGH)
• BKG.L, BMW.DE, MBG.DE, MOS, APA (WATCH)
• PAH3.DE (PASS-biased counter-prediction for IC validation)

Observations (Non-Quantitative) #

Pipeline behavior review:

1. Graham MOS broken for non-USD tickers. GBp/EUR unit mismatches cause nonsensical output (-6349%, -3830%, -2678%, -3003%). Affects UHR.SW, BKG.L, PSN.L, SHEL.L, AKE.PA. Manual IV recomputation in deep analysis corrected. Fix candidate: auto-detect GBp tickers and rescale EPS by 100 before Graham formula. pm/graham/screener.py likely needs currency-aware unit handling.

2. Lynch classifications frequently wrong at top of list. UHR.SW → "fast_grower" should be turnaround. PRU.L → "slow_grower" should be stalwart. PSN.L → "fast_grower" should be cyclical. LLM step under-calibrating — prompt could include historical revenue CAGR bucket.

3. Schilit "0 flags" is the modal outcome (69/322 = 21%) but doesn't differentiate quality. Current scoring gives 100% credit for zero-flags, 85.7% for 1-flag — low discriminating power. Consider adding Altman Z-score or earnings quality score alongside.

4. Reddit divergence yielded exactly 1 signal (PRU.L) — precisely the intended use case. shr-026 false-positive filter worked: no noisy picks from the reddit-hyped tickers (NVDA, BTBD, AIXI) surfaced to the top.

5. Previous cycle candidates (AIG, DD, LEN, TROW, HUM, PHM, BEN) remain in composite top 21 — temporal stability is good (portfolio isn't whipsawing on noise), but also means the pipeline's novelty rate is low. Expected given 14-day window.

Self-Critique #

• What worked: Reddit divergence correctly identified the one contrarian BUY (PRU.L). Deep analysis caught 3 structural-MOS errors (UHR/BKG/PSN data bugs). shr-034 discipline held — no T1 deployments recommended for candidates without AMBER flags.
• What to improve: Pipeline's LLM Lynch classification reliability. Graham MOS unit-conversion bug. Composite score ties at 55.0 for 7 stocks (ranks 5-14) suggest insufficient scoring resolution among middle-of-pack candidates — need finer-grained secondary sorts.

Next Cycle #

• Scheduled: 2026-05-04 (14 days from cycle start)
• First feedback-eligible: 2026-05-03 (prior cycle predictions resolvable)
• Intermediate catalyst dates: MBG Apr 29, BMW May 6, APA May 6, MOS May 5, PAH3 May 13 (Q1 2026 earnings)
• Reconsider PRU.L position sizing after Q1 APE update (expected late Apr / early May)