AI Research TSLA

TSLA daily-return tails vs Gaussian over the last ~3 years

TSLA’s daily returns over the last ~3 years show a clear breakdown of the bell curve: we observed 7 two‑sided |z|≥3 sessions versus 2.03 expected (about 3.45× inflation), and a binomial test gives p = 0.0048 — far less likely than chance. The return series also displays visible heavy tails and positive skew, so extreme moves are both large and asymmetric.

This study resampled minute bars to daily closes, standardized close‑to‑close returns, counted two‑tailed exceedances and measured excess kurtosis. The short thesis is simple: normal‑based daily‑return models materially understate how often TSLA detonates into true 3σ events. The charts, full stats and significance tests that support this finding follow below.

The research question

For TSLA over the past ~3 years, do its daily returns break the bell curve — how many 3-sigma-plus sessions actually occur versus the ~2 a normal distribution predicts over this many trading days, and how heavy are the tails and kurtosis? Thesis: TSLA logs several times more extreme days than Gaussian math allows, so any risk model built on normal returns badly understates how often the stock detonates in either direction.

How this was measured

Resampled TSLA minute bars to daily closes, computed close-to-close returns, and standardized them to z-scores using the in-sample mean and sample standard deviation over the last ~3 years (ending at the latest available date). Counted two-tailed exceedances beyond 2σ and 3σ and compared against normal-theory tail probabilities p=2·(1−Φ(k)). Binomial tests (greater-tail) assess whether observed exceedances are significantly above Gaussian expectations. Excess kurtosis (Fisher) summarizes tail heaviness.

The key numbers

Trading days analyzed
751
2023-06-30 to 2026-06-30
Mean daily return
0.1298%
Std dev of daily return
3.7102%
Excess kurtosis
3.330
excess>0 → heavy tails
Skewness
0.465
Observed 3σ exceedances (two-tailed)
7
Upper=4, Lower=3
Expected 3σ exceedances (Gaussian)
2.03
p=0.0027 per day
Observed/Expected ratio (3σ)
3.45
ratio=3.45 > 1 → heavier-than-normal tail
Binomial p-value (>|3σ| count > Gaussian)
0.0048
p=0.0048 < 0.05 → exceeds Gaussian by chance-unlikely
Observed 2σ exceedances (two-tailed)
32
Upper=20, Lower=12
Expected 2σ exceedances (Gaussian)
34.17
p=0.0455 per day
Observed/Expected ratio (2σ)
0.94
ratio=0.94 ≤ 1 → at/under Gaussian
1st–99th percentile band
18.6669%
1%=-8.59%, 99%=10.08%
Max single-day gain
23.0998%
Max single-day loss
-17.9930%

Reading the numbers

Over 751 trading days TSLA produced 7 two‑sided 3σ sessions versus about 2 expected under a Gaussian (Observed/Expected = 3.45), and the binomial p=0.0048 says that excess is unlikely to be random — extreme days are materially more common than a normal model predicts.

The charts

TSLA standardized daily returns (z-scores)
What this chart says

This histogram plots TSLA daily returns after converting them to z‑scores (n=751, mean centered at 0). What to look at are the far ends: the leftmost return reaches −4.8846 and the rightmost reaches 6.191, so there are very large moves on both sides rather than a tight bell. Those long tails, combined with the positive skew and excess kurtosis reported in the headline stats, show that returns are more extreme and a bit right‑leaning compared with a normal bell curve.

Observed vs Gaussian-expected tail counts
What this chart says

The two bars compare observed counts to Gaussian expectations for |z|≥2 and |z|≥3. For |z|≥2 TSLA had 32 observed events versus ~34.17 expected (so roughly in line), but for |z|≥3 the stock produced 7 observed events versus only ~2.03 expected — about 3.45× more 3σ days. Note the asymmetry in direction (3σ upper=4, lower=3; 2σ upper=20, lower=12) and the binomial p=0.0048 from the headlines, which reinforces that the excess 3σ tail frequency is not just sampling noise.

Tail exceedance summary (two-tailed)

thresholdn_daysp_gaussianexpected_countobserved_countobs_over_expupper_taillower_tailbinom_p_greater
|z| ≥ 27510.045534.17320.9420120.6724
|z| ≥ 37510.00272.0373.45430.0048

Top 10 absolute z-score days (most extreme)

datereturnz_scoretail_sign
2025-04-090.2316.1911
2025-03-10-0.1799-4.8846-1
2024-04-230.15914.25421
2024-04-290.14453.85851
2025-03-240.1323.5231
2025-06-05-0.1273-3.4659-1
2025-04-04-0.1225-3.3375-1
2024-11-110.1092.90251
2024-11-060.10832.88441
2024-10-230.1012.68641

The takeaway

Short answer: yes — TSLA produces far more extreme one-day moves than a Gaussian would predict. Over 751 trading days we saw 7 two‑sided |z|≥3 sessions versus 2.03 expected (about a 3.45× inflation), and a binomial test gives p=0.0048 — only about a 5‑in‑1,000 chance this gap is random. By contrast the |z|≥2 band is not elevated: 32 observed vs 34.17 expected (obs/exp ≈0.94). The distribution is visibly heavy‑tailed (excess kurtosis = 3.33) and skewed to the upside, and the largest single‑day moves reached +23.10% and −17.99%, so extremes are both big and more frequent at the far tail. Practical takeaway: normal‑based daily‑return models materially understate how often TSLA detonates into 3σ‑class moves; the signal for extreme‑tail fatness is statistically strong, even though the absolute count of extremes is modest.

The fine print