Benchmark: Global X China EV & Battery ETF (2845.HK) · daily log returns, 6-month window
(2026-02-13 → 2026-08-11) · Dimson = 1-lag sync-corrected beta · α = realized alpha (stock − β·benchmark),
not annualized · benchmark returned -9.2% over the window.
Static one-time snapshot — not yet on a live datafeed.
name
sector
β raw
β Dimson
se(β)
R²
tot ret
α window
α 63d
6880 Momenta ⚠ 24d
ADAS
1.36
1.40
0.38
0.36
-5%
-6%
n/a
3931 CALB
Battery
1.15
1.22
0.14
0.36
-30%
-25%
-32%
3750 CATL
Battery
1.01
0.78
0.14
0.32
+23%
+30%
+21%
9927 Seres
OEM
0.77
0.98
0.13
0.24
-56%
-74%
-35%
9660 HorizonRobot
ADAS
0.69
0.59
0.16
0.14
-44%
-51%
-15%
9863 Leapmotor
OEM
0.64
0.78
0.16
0.13
-10%
-5%
-9%
1211 BYD
OEM
0.62
0.62
0.12
0.19
-8%
-2%
+1%
2015 Li Auto
OEM
0.58
0.53
0.14
0.13
-32%
-33%
-28%
9868 XPeng
OEM
0.58
0.70
0.15
0.12
-34%
-36%
-19%
2238 GAC
OEM
0.57
0.57
0.10
0.24
-42%
-48%
-8%
9866 NIO
OEM
0.49
0.46
0.17
0.07
-8%
-4%
-15%
7489 Voyah ⚠ 96d
OEM
0.48
0.72
0.19
0.07
-57%
-80%
-62%
2665 Seyond
Lidar
0.43
0.68
0.31
0.02
-76%
-139%
-93%
2525 Hesai
Lidar
0.42
0.65
0.18
0.04
-25%
-26%
-6%
2333 GWM
OEM
0.42
0.28
0.10
0.12
-34%
-37%
-22%
1810 Xiaomi
Diversified
0.41
0.06
0.14
0.07
-27%
-27%
-8%
0175 Geely
OEM
0.41
0.75
0.14
0.07
+7%
+10%
-12%
9973 Chery
OEM
0.21
0.21
0.13
0.02
-10%
-8%
-12%
7687 Eacon ⚠ 24d
ADAS
-0.02
0.57
0.21
0.00
-17%
-19%
n/a
What the data says
Beta ranks exactly as the ETF's composition predicts. Battery names top out
(CALB 1.15, CATL 1.01 — they are the ETF's core weights), OEMs cluster 0.4–0.8,
and the diversified/idiosyncratic names sit lowest (Xiaomi 0.41, Chery 0.21). A coherent,
believable cross-section.
CATL is the lone positive-alpha standout (+30% window, +21% last 63d) — it beat a
falling theme materially. Geely is the only other positive (+10%). Everything else
underperformed its beta-implied return, consistent with a brutal 6-month tape for the sector.
Dimson (sync-corrected) betas diverge sharply for the thin names — Geely 0.41→0.75,
Xiaomi 0.41→0.06, Seyond/Hesai/Voyah all jump ~0.25. Non-synchronous trading is real and
name-specific; the Dimson beta is the more trustworthy of the two.
Caveats
Three names are unreliable — treat as blank. Momenta & Eacon have only 24 returns
(se 0.38 / β −0.02 is noise); Voyah 96. They need more history before any beta is meaningful
(flagged ⚠ above).
The extreme alphas are a low-beta artifact, not signal. Seyond's −139% window alpha
comes from a genuine idiosyncratic collapse (−76%, ~7 → ~2.7) attributed almost entirely to
alpha because its beta is near-zero and R² is 0.02 — the theme barely explains it.
Read big alphas on low-R² names as "this stock did its own thing," not "great/terrible vs the theme."
This window is one regime. Realized alpha here describes a specific down-tape; it's
descriptive, not predictive. Prices are also unadjusted for dividends (minor).
Methodology & why
Benchmark — and why
2845.HK — Global X China EV & Battery ETF tracks the Solactive
China EV & Battery Index, (capped) market-cap weighted with CATL a top holding — so it's
battery-heavy, not equal-weight EV. Because the universe is China-EV names, this thematic
ETF is the right single-ticker benchmark: β measures sensitivity to the EV/battery theme,
and α is the stock-specific excess return over the theme — the stock-selection edge the
signal is trying to capture.
How it's computed
Daily log returns, 6-month window (2026-02-13 → 2026-08-11), each name tail-aligned by
date to the benchmark (recent IPOs use their available overlap).
β raw = OLS slope of stock returns on benchmark returns.
α window = stock cumulative log return − β × benchmark cumulative log return, over the
window. Not annualized.α 63d = same, last 63 trading days (~3-month proxy).
Design decisions — and why
Report horizon alpha, never annualized. Annualizing a daily intercept (×252) over a short,
violent window explodes to a meaningless ±100%. The honest number is realized alpha over the holding
period. For the 3-month horizon a tradeable signal should be scored on forward 63-day
realized alpha — note this snapshot shows realized trailing alpha (descriptive), not the
forward version.
Use a sync-corrected beta. Thin trading biases the plain daily beta. Dimson moves several
names materially, in both directions — up for laggy names (Geely 0.41→0.75, XPeng 0.58→0.70),
down for a few (Xiaomi 0.41→0.06, GWM 0.42→0.28). It reduces sync bias but adds estimation variance,
so part of those swings is noise; weekly returns are a lower-variance alternative for production.
The benchmark is battery-heavy, so read β accordingly. Cap-weighted with CATL/materials
dominant, OEMs land at β 0.2–0.8 with low R² (0.02–0.36) — single-name idiosyncratic vol dominates.
Interpret β<1 as "less battery-exposed," not "defensive." 2845 stays the right single-ticker
choice; blend toward a purer OEM benchmark only if that distinction ever matters.
Full write-up: docs/methodology-alphabeta.md in the repo.