I was wondering whether SABR model (or some of its modifications) is actually used by practionarers.
Also, if one models the FX forward with SABR, would the cocycle condition be satisfied? That is, if $FXF^{ccy1ccy2}$ and $FXF^{ccy2ccy3}$ follow a SABR dynamics, does the same hold for $FXF^{ccy1ccy3}$? And how would the $\alpha, \beta, \rho$ parameters of the third forward read in terms of the ones of the other two? (here, I denote by $FXF^{ccy1ccy2}$ the FX forward rate wrt ccy1 and ccy2).
Assuming $X_t$ and $Y_t$ are diffusion processes and $Z_t=X_t Y_t$, then $$ dZ_t/Z_t = dX_t/X_t+dY_t/Y_t + ... dt $$ So if the diffusive parts of $X_t$, $Y_t$ and $Z_t$ are SABR, then $$ \sigma^Z_t Z_t^{\beta^Z-1} dW^Z_t + ... dt = \sigma^X_t X_t^{\beta^X-1} dW^X_t + \sigma^Y_t Y_t^{\beta^Y-1} dW^Y_t + ... dt $$ and a first necessary condition is $\beta^X = \beta^Y = \beta^Z = 1$.
Next, you have $\sigma^Z_t=\sqrt{(\sigma^X_t)^2 + (\sigma^Y_t)^2 + 2 \rho_{XY}\sigma^X_t \sigma^Y_t}$ and
$$
d\sigma^Z_t/\sigma^Z_t = ((\sigma^X_t)^2 d\sigma^X_t/\sigma^X_t + (\sigma^Y_t)^2 d\sigma^Y_t/\sigma^Y_t+\rho_{XY}\sigma^X_t \sigma^Y_t(d\sigma^X_t/\sigma^X_t+d\sigma^Y_t/\sigma^Y_t))/(\sigma^Z_t)^2 \\
+ ... dt
$$
and another necessary condition is that $d\sigma^X_t/\sigma^X_t = d\sigma^Y_t/\sigma^Y_t$, which in turns implies that if $X_t$, $Y_t$ and $Z_t$ are SABR, then $\alpha^X=\alpha^Y=\alpha^Z$, $\rho_{XY}$ is constant, and $\sigma^X_t/\sigma^X_0 = \sigma^Y_t/\sigma^Y_0 = \sigma^Z_t/\sigma^Z_0$ (all three stochastic volatilities are driven by the same brownian motion and have the same volvol).
The SABR parameters $\rho^Z$ is then computed as
$$
\rho^Z = (\sigma^X_0 \rho^X + \sigma^Y_0 \rho^Y)/\sqrt{(\sigma^X_0)^2 + (\sigma^Y_0)^2 + 2 \rho_{XY}\sigma^X_0 \sigma^Y_0}
$$
and an additional necessary condition is that $|\rho^Z| \leq 1$.
It's easily checked that put together, these necessary conditions are also sufficient, but it's all rather restrictive.
That being said, SABR is usually not used as a model but rather as a sparse parameterization of the implied volatility smile, so there is nothing wrong with fitting distinct SABR to each of the currency pair in the triangle, although the vanna-volga parameterization might be more popular for FX smiles.
