Raw-intercept alphas absorbed the T-bill yield on uninvested/levered
portions (582 well-fitted funds >2%/yr off; sum-of-betas polluted by
level-matching). Now fund AND sleeves are netted against BIL daily
total return before every regression; a cash position contributes
exactly zero.
- decompose: rf_series()/excess(); shv+bil dropped from regressors
(~0 columns in excess space); FULL_WINDOW -> 2007-06-01 (BIL
inception; mixing raw pre-2007 with excess breaks the fit).
- factors: same excess treatment; shv out of DRIVERS.
- CASH axis redefined: alpha/cash_yield -> net cash position = 1 -
sum(betas) (label 'cash (net posn)').
- CANDIDATE list 250 -> 11: the old list was mostly under-invested
funds whose 'alpha' was cash yield, not skill.
- refback.py: per-fund fitted reference (forward-selected sleeves)
stored as ref_5y/ref_full in search_all.json; app alpha-search
table gains a 'reference (5y)' column - the answer to 'what is
alpha computed against' (the fund's OWN fitted sleeve mix, not one
index).
- App captions updated; raw-alpha-era results backed up as
*_rawalpha.json (not deleted).
factors.py: full OLS of all 2,384 funds on a 35-driver basis
(overinclusive, no portfolio-corr screening - corr is a replacement
signal, not a rejection). v1's 21 + lqd/hyg/prefs/emb/tip/shy/vtv/
8 sectors/CTA/commodities. Basis fixes: drop vea/vug (dupes of
efa/qqq), drop finux (TERMINATED 2017 - silently zeroed the
complete-case mask; v1 forward selection never hit this), drop bil
(shv/bil near-null -> offsetting shv+64/bil-54 noise fits),
residualize vblix on ivv+tlt (pure vol axis), ridge 0.02.
cluster.py: hierarchical tree saved but fixed-k cuts degenerate
(most funds are blends -> one 2250-fund blob); k-means++ (deterministic)
is the useful grouping, re-run live in-app for any k.
app: Return-driver clusters expander (k slider 10-60, summary table,
member table sorted by alpha-t). Findings at k=30 in RESEARCH.md.
