A small network cuts Heston barrier-call RMSE by about 12 while seeing only two prices and one spacing measure. At every grid point, it receives the coarse-mesh price and the refined-mesh price at the same node, then returns an estimate close to the reference-mesh price. Those two values make up the full input on the uniform Black-Scholes grids. Because the Heston grid is non-uniform, the network also receives a local mesh-spacing measure. In the down-and-out barrier call test, RMSE drops from 7.4346e-3 to 6.2839e-4. Coordinates, strike, volatility, barrier level and boundary data are all withheld.
We could not test the method on our own data. Our option history covers end-of-day listed US equity and ETF contracts, while no market price series exists for a three-asset cash-or-nothing digital or a Heston knock-out. Listed vanillas would change both the payoff and the validation problem, leaving the reported error reductions inapplicable.
Two values, then a spacing measure
Cobb, Hao and Liu construct three nested meshes for each pricing PDE. The coarse mesh comes first, the refined mesh doubles its resolution, and the reference mesh has the highest resolution. Each coarse node coincides with one on the refined grid. A collocation point therefore supplies two computed values for the same solution, drawn from grids a factor of two apart and converging toward the same answer.
On the uniform Black-Scholes grids, those two values are everything the network sees. The Heston grid, built following in 't Hout and Foulon, contributes a third number: h_local = sqrt((h_left + h_right)(h_up + h_down)). This local mesh-spacing measure uses the distances to the four neighbouring nodes. The reference value at the point becomes the target.
The paper says explicitly that the network receives no spatio-temporal coordinates, no PDE hyperparameters such as volatility, rate or mean reversion, and no metadata for boundary or initial conditions. Its argument is that the pair of converging solution values already contains this information. I read the mechanism as a nonlinear learned version of Richardson extrapolation: the gap between coarse and refined values at a shared node samples the leading local discretization error, and the network learns the coefficient. The authors give another motivation. These parabolic solutions are monotone in each asset price, they write, allowing a single-point patch to improve results where their earlier hyperbolic work required 8 or 18 inputs.
The networks remain tiny. They use two hidden layers, with 15 neurons for 1D, 20 for 2D and 3D, and 32 for Heston. Training uses ReLU, Adam at 1e-3, and 1500 to 2000 epochs. Black-Scholes is solved with first-order ADI, while Heston uses Modified Craig-Sneyd. Every result comes from synthetic PDE solves or a closed-form formula. The paper contains no market data.
How much error disappears?
The experiments span two option classes across four settings. The first is a cash-or-nothing call under N-asset Black-Scholes, with K=100, C=100, S_max=300 and T=1. Cases N=1,2,3 are compared with the closed-form solution. The second is a single-asset Heston down-and-out barrier call, using K=100, T=1, S_max=800 and v_max=5. Since no closed form exists, its benchmark is the 200x100 reference mesh.
For 1D, the test-set reduction ranges from 2.5x to 3.2x. At training gap g=4, the refined result is 2.019 and NNLCI reaches 0.623; g means retaining every g-th node along each parameter axis. In 2D at g=2, the refined value of 9.206 falls to 1.938 with NNLCI, about 4.8x. The 3D refined test RMSE spans 10.7 to 11.6, then drops to 2.32 to 3.41.
Heston produces the largest improvement. Its non-uniform grid and knock-out feature, where the option dies when spot touches the barrier, make it the hardest case. The reduction is roughly 12x for g=2 and 14x for g=4. Those tables also extend beyond both ends of the abstract's 4-12x range. The 1D test reductions of 2.5x to 3.2x fall below 4, while Heston at g=4 reaches about 14x, above 12.
The hardest case delivers the largest reduction.
Data efficiency is the genuinely interesting result. In 2D at g=8, training uses 2^4 = 16 of 6,561 parameter tuples, about 0.24%. Test RMSE still reaches 2.312, compared with 9.236 for the refined solver. The Heston g=4 experiment uses 1 x 2^5 = 32 of 12,500 tuples and gives the stronger of the two Heston outcomes.
More training does not consistently lower test error. In 1D, 0.799 at g=2 with 81 training pairs is worse than 0.680 at g=16 with 4 pairs. The 3D results behave similarly: the best test RMSE comes from m=2 training examples out of 20, while m=8 is worst. The authors caution that m=2 leaves a test set of 18 random tuples, and that "individual m-comparisons should be read with that variability in mind."
Exact formula versus finer mesh
Black-Scholes errors are calculated against the exact formula. Heston errors instead use a finer mesh as the benchmark, so any discretization bias shared by the refined and reference grids cancels from the reported Heston number. The 7.4e-3 figure measures distance between two approximations, with the network trained against a target carrying its own error. The 12x Heston reduction and the 3.2x Black-Scholes reduction therefore measure different quantities.
Absolute Black-Scholes baseline errors are weak. Across 2D and 3D, refined-mesh test RMSE ranges from 9.21 to 11.65 on a payout of C=100, roughly a 10% error. In 1D, the range is 1.9996 to 2.0241, about 2% of the payout. A first-order scheme produces those figures with 21 and 41 nodes per direction and 21 and 41 time steps. Reducing the 3D error to 2.3 has value. The remaining error is still 2.3 on a 100 payout.
The missing clock
The abstract closes by claiming that the results "demonstrate that NNLCI significantly reduces computational requirements for high-dimensional problems in real-time options trading and risk management, offering low training costs and strong generalization ability." Evidence supports the final two clauses. The Heston g=4 run trains on 32 tuples, which is a low training cost, and its generalization test holds out 99.74% of the parameter grid. For 2D at g=8, the held-out share reaches 99.76%.
Computational requirements receive no direct measurement. The Introduction gives a qualitative mechanism: after training, the network maps a coarse solution and a refined solution to an accurate value at far lower cost than solving the reference mesh during inference. Mesh counts supply the paper's sole cost argument, which I bound below.
No wall-clock time or FLOP count is reported for either the solvers or the network. The paper makes no comparison with Monte Carlo, PINN or DeepONet baselines. It also leaves unanswered what a finer mesh could have achieved with the same compute.
The reported grids allow a bound on the saving. Heston inference requires coarse and refined solves of 50x25 plus 100x50 cells, against a 200x100 reference. My arithmetic from those grid sizes gives about 3.2x fewer cells before allowing for the per-cell cost of ADI sweeps or network passes. That 3.2x sets the ceiling for the efficiency claim. It applies only to Heston, the sole experiment requiring a reference-mesh computation. Black-Scholes used the closed form and had no reference solve.
The authors acknowledge the theoretical gap. A local coarse patch has no guarantee of uniquely determining the reference solution, because distinct global solutions can share local patterns. They argue that local, converging patches of coarse solutions substantially reduce this risk, though the paper does not measure the reduction. The claim remains asserted.
Earlier dissertation work in reference [6] is also discussed. NNLCI failed to converge beyond a short time horizon for the double pendulum and three-body problems because of sensitive dependence on initial conditions. The authors therefore limit this paper to non-chaotic parabolic PDEs. Larger patches, 3x3 and 3x3x3, appear as future work. Only the single-point patch is tested here, using two inputs on the uniform grids and three on the Heston mesh. I also did not find a validation split separate from the test set. Network widths and epoch counts change between experiments without a stated selection rule.
A table of seconds would change my view: reference-mesh solve, coarse plus refined plus network inference, and Monte Carlo at matched accuracy for the 3-asset case where dimensionality carries the real cost. Until such a comparison appears, the paper stands as a well-executed numerical error-correction result with very little training data. Its real-time trading claim remains a sentence.