AI Research MSFT

MSFT leverage-effect test: do losses drive higher next-day intraday volatility than equal gains? (≈3y window)

MSFT shows a measurable leverage effect over the past ~3 years: a same-sized decline tends to be followed by a larger next-day intraday range than an equal-sized gain. In an OLS model on 729 trading days the signed-return coefficient is c = −0.48346 (p = 0.00597), which translates to about a 0.97% bigger next-day high–low range after a −1% close. That result is statistically robust but modest in scale.

The study resampled minute bars to daily OHLC, regressed next-day (high−low)/close on |r_t| and r_t with Newey–West errors, and also ran bucketed comparisons. The R-squared is low (0.014), so sign/magnitude explain only a small slice of range variability; the detailed charts, tests and caveats are below.

The research question

For MSFT over the past ~3 years, does a down day spike the next session's volatility more than an up day of the same size — the asymmetric 'leverage effect' where fear drives volatility harder than greed? Thesis: a given-magnitude decline is followed by a meaningfully larger next-day high-low range than an equal-magnitude gain, so volatility reacts to losses rather than to moves in general.

How this was measured

Resampled MSFT minute bars to daily OHLCV over the last ~3 years. Computed daily close-to-close return r_t and next-day intraday volatility proxy as (high−low)/close on day t+1, aligned to r_t. Tested asymmetry two ways: (1) OLS regression vol_{t+1} = a + b·|r_t| + c·r_t with HAC (Newey-West, 5 lags) standard errors; a negative, significant c indicates that, holding magnitude fixed, losses precede higher next-day volatility than gains. (2) Bucketed days by absolute return quintiles and compared mean next-day ranges for negative vs positive days within each bin via Welch's t-test.

The key numbers

Trading days analyzed
729
2023-07-03 to 2026-05-28
Mean next-day high–low range
11.4232%
Vol proxy = (high−low)/close on t+1
Intercept (a)
10.8560%
Baseline next-day vol when r_t = 0
Magnitude coefficient (b on |r_t|)
0.54217
Sensitivity of next-day vol to move size
Asymmetry coefficient (c on r_t)
-0.48346
c=-0.48346 < 0 → losses raise vol more than equal gains
p-value for c
0.0060
Two-sided; p=0.0060 < 0.05 → asymmetry detected
R-squared (OLS)
0.014
Implied Δ vol (neg−pos) after ±1% move
0.9669%
Computed as −2·c·0.01 from the regression
Implied Δ vol (neg−pos) after ±2% move
1.9338%
Computed as −2·c·0.02 from the regression
Top-magnitude bin Δ (neg−pos) [1.72%–11.60%]
1.5100%
N_pos=74, N_neg=72
Top-bin Welch p-value
0.1861
p=0.1861 ≥ 0.05 → no clear asymmetry at largest moves

Reading the numbers

Across 729 trading days the average next‑day high–low range was 11.42%. The regression finds a negative asymmetry coefficient c = -0.4834568808721459 (p = 0.005970546843358322), implying about a 0.009669137617442917 (0.9669%) larger next‑day range after a 1% drop than after a 1% rise, though R² is only 0.014016229125365376.

The charts

Mean next-day high–low range by |return| bin and sign
What this chart says

This bar chart splits mean next‑day (H−L)/C by today's absolute return bin and by sign. Look at the 1.04–1.72% bin where negative days lead to a mean next‑day range of 0.1272 versus 0.0974 after positive days — the biggest gap in the chart. The top magnitude bin (1.72–11.60%) also shows higher next‑day range after losses (0.1264 vs 0.1113; neg−pos = 0.0151 noted above). The pattern of losses producing slightly larger next‑day ranges holds in most bins, but the differences are numerically modest.

MSFT signed return (t) vs next-day high–low range (t+1)
What this chart says

The scatter plots each day's signed return (x from -0.0631 to 0.116, mean 0.0005) against the next‑day range (y from 0.0095 to 0.6716, mean 0.1142) for n=729. The cloud is wide with no tight slope, which matches the low OLS R² = 0.014016229125365376: most of the variation in next‑day range is noise. So although the regression picks up a statistically significant negative asymmetry, the scatter shows the effect is small relative to day‑to‑day dispersion and driven partly by a few large outliers.

Per-bin next-day volatility summary (pos vs neg)

abs_ret_binN_posN_negmean_vol_next_posmean_vol_next_negdiff_neg_minus_poswelch_p
-0.10%–0.29%66790.10080.10940.00860.5007
0.29%–0.61%84620.10110.1120.01090.3506
0.61%–1.04%73720.12990.13030.00040.9782
1.04%–1.72%78680.09740.12720.02980.0085
1.72%–11.60%74720.11130.12640.01510.1861

The takeaway

Yes — over the last ~3 years MSFT shows a measurable leverage effect: losses of a given size tend to be followed by larger next-day high–low ranges than equal-sized gains. The signed-return coefficient in the OLS model is c = -0.48346 with p = 0.00597 across 729 trading days, which translates to about a 0.97% bigger next-day range after a −1% day (0.00967) and about 1.93% after a ±2% move (0.01934). That result is statistically robust but small in explanatory power: the regression R-squared is only 0.014, so the sign/magnitude story explains about 1.4% of day-to-day range variability. The bucketed tests show the asymmetry is clearest in the mid-large bin (1.04%–1.72%) where neg–pos gap = 2.98% (p = 0.0085), while the largest-move bin (1.72%–11.60%) has a 1.51% gap that is not significant (p = 0.1861). Bottom line: there is a real, statistically significant lean that losses spike next-day intraday volatility more than gains, but it’s modest in size and not uniform across the biggest moves — treat it as a useful tilt, not a dominant driver of volatility.

The fine print