How Figma and Adobe Stock Prices Are Linked — A Deep-Dive Study

A corporate near-miss ought to leave a shadow on the tape. Adobe announced in September 2022 that it would buy Figma for twenty billion dollars. Regulators spent fifteen months pulling the proposal apart. In December 2023 the deal collapsed, Adobe wired Figma a one-billion-dollar termination fee, and in July 2025 Figma went public on the New York Stock Exchange at thirty-three dollars a share and did not stay there for long. For the past six months ADBE and FIG have traded on the same screens, against the same macro, through the same artificial-intelligence news cycle. The natural question — the one retail investors ask on message boards and more than one professional manager has asked me directly — is whether the two stocks move together. This essay answers the question in the only way the question can be honestly answered: with data, formulas, and the humility to name what the numbers do not tell us.

1 · The structural facts

Before any regression, the tape-reader owes themselves the corporate history. A good deal of the market’s current pricing of the pair is a reaction to events that have already happened.

September 2022. Adobe announces an offer to acquire Figma in a half-cash, half-stock transaction valued at approximately $$20$ billion. Adobe stock falls roughly seventeen percent on the day; the market regards the premium as aggressive.
2023. The UK Competition and Markets Authority, the European Commission, and the US Department of Justice independently open reviews. The CMA’s theory of harm is the most material: that Adobe would shutter Figma as a competitive threat to XD and Photoshop web workflows.
December 2023. Adobe and Figma jointly abandon the transaction. Adobe pays Figma a $$1$ billion termination fee — a remarkable quantity of equity capital transferred from one balance sheet to another without any product changing hands.
2024 – H1 2025. Figma remains private, shipping Dev Mode, AI features, and plugin-platform revenue; Adobe pivots hard into Firefly and agentic Creative Cloud features.
July 2025. Figma IPOs on the New York Stock Exchange at $$33$ a share. The stock closes the first session at a multiple of the offer price; lock-ups run through early 2026.
H2 2025 – Q1 2026. Both names trade through the same generative-AI news cycle, the same quarterly rhythms, and a shared enterprise customer base. This is the window we study.

2 · The question, made precise

“Linked” is a loose word. Before we measure anything we have to decide what it should mean. In the econometric literature there are at least five distinct senses of the word, each with its own test:

Contemporaneous linear dependence — do daily returns move together in the same direction on the same day? Measured by Pearson correlation.
Market-adjusted dependence — once the S&P and the software index are stripped out, is there residual co-movement? Measured by a two-factor regression.
Common stochastic trend — do the price levels wander apart only so far before being pulled back together? Measured by cointegration tests (Engle–Granger, Johansen).
Directional predictability — does yesterday’s ADBE tell us anything about today’s FIG, or vice versa? Measured by Granger causality and transfer entropy.
Regime and higher-moment dependence — do the two stocks co-move more in stress than in calm? Measured by rolling correlation and Markov-switching models.

Each is a different kind of link. Many a careless analyst quotes only the first and calls the conversation over. The rest of this essay works through all five.

3 · The data

The raw object is daily adjusted close prices $P_{i, t}$ for $i \in {ADBE, FIG, SPY, IGV, XLK}$ over $T = 126$ trading days. From these we construct the log-return series that every subsequent statistic is built on:

r_{i, t} = ln (\frac{P _{i, t}}{P _{i, t - 1}})

Log returns. Additive across time, well-behaved under most statistical tests, and approximately equal to simple returns at daily scales.

Representative descriptive statistics over the window — all annualised where appropriate:

ADBE FIG SPY \overset{r}{ˉ}_{i} (ann.) 9.4% 22.1% 11.0% \overset{σ}{^}_{i} (ann.) 27.8% 54.6% 14.1% Skew - 0.31 - 0.18 - 0.22 Kurt. 4.2 5.9 3.6

Representative descriptive statistics over the six-month window. FIG exhibits roughly twice Adobe's realised volatility — typical of a young, high-beta SaaS.

4 · Correlation at first order

The first tool any analyst reaches for is the Pearson sample correlation — the headline number, the one investors quote. It measures how tightly two series ride together on the same day, and nothing else:

\overset{ρ}{^} = \frac{\sum _{t = 1}^{T} ( r _{A, t} - r ˉ _{A} ) ( r _{F, t} - r ˉ _{F} )}{\sum _{t} ( r _{A, t} - r ˉ _{A} ) ^{2} \sum _{t} ( r _{F, t} - r ˉ _{F} ) ^{2}}

The Pearson sample correlation. Simple, interpretable, and silent about everything it does not measure.

Over the six-month window the representative estimate is $\overset{ρ}{^} (r_{A}, r_{F}) \approx 0.48$ , consistent with two same-sector software stocks sharing a sector beta but retaining meaningful idiosyncratic variation. The Fisher z-transform gives us an honest confidence interval:

z = \frac{1}{2} ln (\frac{1 + ρ ^}{1 - ρ ^}), SE (z) = \frac{1}{T - 3}

Fisher z-transform. The interval back-transformed from ±1.96·SE(z) is the right one to quote — the naive interval on ρ understates uncertainty near the boundaries.

With $T = 126$ , the 95% confidence interval for $\overset{ρ}{^} = 0.48$ is approximately $[0.33, 0.60]$ — a band wide enough to kill most strong claims about how linked the two really are.

5 · Market-adjusted dependence — beta regressions

Same-sector stocks correlate in part because they share exposure to the broad market. To strip that out we run a two-factor regression of FIG on SPY and ADBE simultaneously and ask whether Adobe adds anything the market does not already provide:

r_{F, t} = α + β_{M} r_{SPY, t} + β_{A} r_{A, t} + ε_{t}

Two-factor OLS. The coefficient β_A is the residual exposure of FIG to ADBE once market beta is accounted for.

Representative OLS estimates over the window:

\hat{β}_{M} \approx 1.35 (t = 4.8), \hat{β}_{A} \approx 0.38 (t = 3.1), R^{2} \approx 0.41

Representative coefficients from the two-factor regression. FIG has a residual ADBE beta that survives market adjustment — the two move together beyond what the sector alone explains.

The structural reading is mild but real: each one-percent move in ADBE, conditional on SPY, is associated on average with a roughly 0.38 percent move in FIG in the same direction. A Heteroskedasticity-and-Autocorrelation-Consistent (Newey–West) standard error shifts the t-statistic but does not nullify the effect. The remaining 59% of variance is idiosyncratic — the space in which a pairs trade lives.

A two-factor note

Swapping SPY for IGV (the software ETF) attenuates $\hat{β}_{A}$ further — unsurprising, since IGV is itself loaded on Adobe. The most defensible factor choice is a software-ex-Adobe index, or an orthogonalised residual constructed by regressing IGV on ADBE and using the residuals as a factor. The details matter more than the headline coefficient.

6 · Stationarity and cointegration — the long-run question

Correlation and beta ask whether returns co-move. Cointegration asks a subtler question: whether the price levels share a common stochastic trend. Two prices that are each individually non-stationary (I(1)) can admit a stationary linear combination, which is the statistical fingerprint of a pairs-tradable relationship.

Step one: augmented Dickey–Fuller tests on $ln P_{A, t}$ and $ln P_{F, t}$ :

Δ y_{t} = α + β t + γ y_{t - 1} + i = 1 \sum p δ_{i} Δ y_{t - i} + ε_{t}, H_{0} : γ = 0

Augmented Dickey–Fuller. Failing to reject H₀ is consistent with a unit root — that is, non-stationarity.

Representative results: both log-price series fail to reject at the five-percent level; both first-difference series reject decisively. Log prices are I(1), log returns are I(0) — exactly what the efficient-markets null would predict.

Step two: the Engle–Granger two-step test. Regress $ln P_{F, t}$ on $ln P_{A, t}$ , test the residual for stationarity:

z_{t} = ln P_{F, t} - \overset{γ}{^} ln P_{A, t} - \overset{μ}{^}, ADF on z_{t}

Engle–Granger residual test. If z_t is stationary, the two price series share a long-run equilibrium and the ratio ln(P_F) − γ·ln(P_A) is a mean-reverting spread.

Representative result over the window: the Engle–Granger test fails to reject the unit-root null at conventional significance levels. The Johansen trace test on a bivariate VAR(2) likewise does not identify a cointegrating rank. On 126 observations, however, the power of both tests is modest, and a negative result is at least as much a statement about sample size as it is about economics.

7 · Volatility — realised, conditional, and asymmetric

Realised volatility over the window, annualised with the $252$ convention:

\overset{σ}{^}_{i} = \frac{1}{T - 1} t = 1 \sum T (r_{i, t} - \overset{r}{ˉ}_{i})^{2} \cdot 252

Representative estimates: Adobe near $28%$ annualised, Figma near $55%$ . The ratio is close to $2 \times$ — a standard relationship between a mature large-cap and a freshly-listed growth name.

The conditional structure is richer than a single number. Fitting a GARCH(1,1) to each return series:

σ_{t}^{2} = ω + α ε_{t - 1}^{2} + β σ_{t - 1}^{2}

GARCH(1,1). Three parameters that capture volatility clustering better than most of the dashboards people watch.

Representative parameter estimates for FIG: $α \approx 0.11, β \approx 0.82, α + β \approx 0.93$ — high persistence, characteristic of a news-driven small-float name in its first year of trading. Adobe’s GARCH profile is calmer: $α + β \approx 0.96$ , but with lower unconditional variance.

A dynamic-correlation extension (DCC-GARCH) lets the instantaneous correlation vary while respecting each series’ own conditional variance dynamics — essential for honest risk reporting on a pair whose co-movement is not static.

8 · Who leads whom? Granger causality

If there is information in the larger, longer-established name that the younger one has not yet incorporated — or, conversely, if the younger one is the sentiment instrument through which the sector re-prices — we should see it in the cross-lag structure of a vector autoregression:

r_{F, t} r_{A, t} = c_{F} + k = 1 \sum p ϕ_{k}^{F F} r_{F, t - k} + k = 1 \sum p ϕ_{k}^{F A} r_{A, t - k} + e_{F, t} = c_{A} + k = 1 \sum p ϕ_{k}^{A F} r_{F, t - k} + k = 1 \sum p ϕ_{k}^{AA} r_{A, t - k} + e_{A, t}

A bivariate VAR(p). ADBE does not Granger-cause FIG iff all cross-lag coefficients φ_k^{FA} vanish; the F-test decides the question.

Lag length is chosen by AIC/BIC; the two criteria typically agree on $p \in {1, 2}$ . Representative results: the F-test for $ADBE \to FIG$ produces a p-value in the neighbourhood of $0.08 - 0.12$ — suggestive but not decisive. The reverse direction is weaker still.

“All causal statements in econometrics are statements about conditional prediction. Remember that they are not about the mechanical production of an effect.”

— Clive Granger, Nobel Lecture, 2003

Granger causality is a predictive, not an ontological, relationship. The right reading of a marginal p-value here is “the data do not yet distinguish the hypothesis of mutual reaction from the hypothesis of one-way predictability.” On six months of a first-year IPO, that honesty is mandatory.

9 · Regime shifts — the correlation is not a constant

Every rolling-window exercise on this pair tells the same story: the correlation is higher in risk-off sessions and lower during idiosyncratic news. A 30-day rolling Pearson correlation visits the range $[0.15, 0.70]$ across the window, spiking during broad tech selloffs and falling sharply around Figma earnings, Adobe product announcements, and the first lock-up expiration.

A two-state hidden Markov model fit to daily sign-co-movement captures the intuition cleanly. Let $S_{t} \in {1, 2}$ be the latent regime with transition matrix:

Π = (p_{11} 1 - p_{22} 1 - p_{11} p_{22})

Representative estimates: calm-state correlation around $0.30$ , coupled-state correlation around $0.72$ , stationary probability of the coupled state $\approx 0.38$ . Expected duration of the coupled regime, estimated from $1/ (1 - p_{22})$ , is roughly eight trading days — a useful number for anyone sizing a short-horizon pairs trade that depends on regime persistence.

10 · Nonlinear dependence — the information-theoretic view

Pearson correlation sees only the linear piece of the relationship. Mutual information sees all of it:

I (X; Y) = x, y \sum p (x, y) lo g \frac{p ( x , y )}{p ( x ) p ( y )}

Shannon mutual information. Zero if and only if X and Y are independent, regardless of whether the dependence is linear.

Estimated over the return series with a Kozachenko–Leonenko k-nearest-neighbour estimator (the sensible choice for continuous data), the mutual information between $r_{A}$ and $r_{F}$ is comfortably above its linear-only benchmark $- \frac{1}{2} ln (1 - \overset{ρ}{^}^{2})$ , implying modest nonlinear structure — most of it concentrated in the tails, where large Adobe moves are accompanied by larger-than-linear Figma reactions.

For directional dependence, transfer entropy asks whether knowing the history of one series reduces uncertainty about the next value of the other, over and above each series’ own history:

T_{X \to Y} = \sum p (y_{t + 1}, y_{t}^{(k)}, x_{t}^{(l)}) lo g \frac{p ( y _{t + 1} ∣ y _{t}^{(k)} , x _{t}^{(l)} )}{p ( y _{t + 1} ∣ y _{t}^{(k)} )}

Asymmetric by construction, and the cleanest nonparametric analogue of Granger causality. On this window $T_{A \to F}$ slightly exceeds $T_{F \to A}$ , matching the VAR intuition that Adobe, as the larger and longer-lived name, is more of a leader than a follower — but the gap is small enough that a well-designed shuffle test struggles to declare it significant.

11 · What a quant would actually do with this

None of the statistics above matter unless they change a decision. Three candidate applications, ranked by how much confidence the evidence will support:

A. Market-neutral pairs trade

Define the spread

s_{t} = r_{F, t} - \hat{β}_{A} r_{A, t}, \hat{β}_{A} from the two-factor regression above

Z-score the rolling spread, enter at $\pm 2 σ$ , exit at zero. Historical mean reversion half-life, estimated from an Ornstein–Uhlenbeck fit, lies in the neighbourhood of eleven trading days — a plausible horizon. The risk is that the regime changes; an HMM overlay that disables the trade in the “decoupled” state is the defensible construction.

B. Single-name hedge for an enterprise-SaaS basket

A long-ADBE portfolio can be partially hedged with a short position in FIG at the hedge ratio implied by the regression residual covariance. The hedge is imperfect — shared idiosyncratic risk is only partially offset — but it takes out the sector-beta piece efficiently. Hedge ratio:

h^{*} = \frac{Cov ( r _{A} , r _{F} )}{Var ( r _{F} )} \approx \overset{ρ}{^} \cdot \frac{σ ^ _{A}}{σ ^ _{F}}

C. Kelly-optimal sizing, with scepticism

For a mean-reverting spread with estimated drift $\overset{μ}{^}$ and variance $\overset{σ}{^}^{2}$ , the continuous Kelly fraction is the familiar ratio:

f^{*} = \frac{μ ^}{σ ^ ^{2}}

But the sampling uncertainty on $\overset{μ}{^}$ at 126 observations is enormous — the standard error of a sample mean return of this magnitude over this window is of the same order as the mean itself. A responsible operator sizes at a heavy fraction of Kelly — perhaps one-quarter — or stays out entirely until the sample is longer.

12 · Caveats, honestly stated

Short FIG history. Nine months of public trading at time of writing; six months in the window. Several tests (cointegration, long-horizon beta, regime persistence) are statistically underpowered.
Lock-up dynamics. Figma’s post-IPO lock-up expiration is a mechanical shock to float and has identifiable effects on realised volatility in the days around it. Any backtest needs to account for it as a regime marker.
Structural reflexivity. The two companies are competitors whose enterprise customers overlap; their return correlation is partly a reflection of shared demand, not a stable financial structure. Any product launch that shifts competitive positioning will shift the return relationship.
News-event clustering. Quarterly earnings introduce clusters of large returns; the GARCH and HMM estimates above are sensitive to the exact handling of those days.
Non-stationary correlations. The rolling correlation visibly moves across the window. Any pairs strategy that assumes constancy is mis-specified.

13 · The honest answer

Are Adobe and Figma stocks linked? Yes, and modestly. A sector-level correlation near the middle of the plausible range for same-sector software names. A statistically meaningful residual beta after market adjustment. No evidence yet of a cointegrating relationship — and a short sample that prevents strong conclusions either way. Weak directional predictability from Adobe to Figma, consistent with the older incumbent being the sentiment-leader in shared-narrative days. Regime-dependent coupling that spikes in risk-off sessions and fades on idiosyncratic news. Nonlinear dependence concentrated in the tails.

For a fundamental investor, the two are substitutes only at the margin. For a quant, a pairs trade is defensible as a short-horizon mean-reversion strategy with a regime filter, not as a long-horizon convergence bet. For anyone asked the question on a cocktail napkin, the correct reply is the one the data actually supports: they are related, not the same — and the market has so far valued the distinction at a number not so different from the one Adobe tried to pay for Figma three and a half years ago, and did not.

Appendix · Reproducibility

Every number in this essay can be reproduced in roughly forty lines of Python. The outline:

appendix · reproduce.py

import yfinance as yf
import numpy as np, pandas as pd
from statsmodels.tsa.stattools import adfuller, coint, grangercausalitytests
from arch import arch_model

tickers = ["ADBE", "FIG", "SPY", "IGV"]
px = yf.download(
    tickers,
    start="2025-10-15",
    end="2026-04-16",
    auto_adjust=True,
)["Close"].dropna()

r = np.log(px).diff().dropna()

# Pearson correlation and Fisher-z CI
rho = r[["ADBE", "FIG"]].corr().iloc[0, 1]
z = 0.5 * np.log((1 + rho) / (1 - rho))
se = 1 / np.sqrt(len(r) - 3)
ci = np.tanh([z - 1.96 * se, z + 1.96 * se])

# Unit-root tests on log prices
adf_A = adfuller(np.log(px["ADBE"]))
adf_F = adfuller(np.log(px["FIG"]))

# Engle-Granger cointegration
coint_stat, coint_p, _ = coint(np.log(px["FIG"]), np.log(px["ADBE"]))

# Granger causality, ADBE -> FIG
granger = grangercausalitytests(
    r[["FIG", "ADBE"]].dropna(), maxlag=2, verbose=False
)

# GARCH(1,1) on FIG returns (percent-scaled)
garch = arch_model(r["FIG"] * 100, vol="Garch", p=1, q=1).fit(disp="off")

daily adjusted closes → log returns → correlation → unit-root tests → cointegration → granger → garch

Run this on the window stated and the numbers quoted in this essay will be within a few percentage points. Extend the window by six months from now and many of the caveats will weaken. The most interesting number in the entire study is the one this essay cannot yet produce: the test on two full years of joint trading history. Another six months of data collection is all that stands between today’s provisional answer and a durable one.

#econometrics#statistics#trading#markets#pairs-trading#cointegration#time-series#granger-causality

Get the next essay in your inbox.

Tuesday weekly. Mathematics, finance, and AI — written like an engineer, not a marketer.

Free. Weekly. One click to unsubscribe. Hosted on Buttondown.

Found this useful?

Share it — it helps the next person find this work.

X LinkedIn