Table 7 Averaged performance of the backward pass of the Heston–CaNN, calibrating 3 parameters on a CPU (Intel i5, 3.33 GHz with cache size 4 MB) and on a GPU (NVIDIA Tesla P100), over \(3125 \times 5\) (random seeds) test cases, where †stands for CaNN estimated value, and ∗ stands for the true value, with \(\mbox{MJ}=J(\varTheta )/N\)
From: A neural network-based framework for financial model calibration
Absolute deviation from \(\varTheta ^{*}\) | Error measure | Computational cost | |||
|---|---|---|---|---|---|
\(|\bar{v}^{\dagger }-\bar{v}^{*}|\) | 1.60 × 10−3 | J(Θ) | 1.45 × 10−6 | CPU time (seconds) | 0.29 |
\(|\gamma ^{\dagger }-\gamma ^{*}|\) | 1.79 × 10−2 | MJ | 4.14 × 10−8 | GPU time (seconds) | 0.15 |
\(|\rho ^{\dagger }-\rho ^{*}|\) | 2.44 × 10−2 | Data points | 35 | Function evaluations | 59,221 |