Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Neural MJD: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction
Authors: Yuanpei Gao, Qi Yan, Yan Leng, Renjie Liao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on both synthetic and real-world datasets demonstrate that Neural MJD consistently outperforms stateof-the-art deep learning and statistical learning methods. Our code is available at https://github.com/DSL-Lab/neural-MJD. |
| Researcher Affiliation | Academia | Yuanpei Gao1,2 EMAIL Qi Yan1,2 EMAIL Yan Leng3 EMAIL Renjie Liao1,2,4 EMAIL 1University of British Columbia; 2Vector Institute 3University of Texas at Austin; 4Canada CIFAR AI Chair |
| Pseudocode | Yes | Algorithm 1 Neural MJD Training 1: repeat 2: (C, S1:Tf ) Dtrain, with C = [S Tp:0, C] 3: {µτ, στ, λτ, ντ, γτ} Tf τ=1 fθ(C) 4: ˆS0 S0 5: for τ = 1, , Tf do 6: ψτ ln P(ln Sτ | Sτ 1 = ˆSτ 1, C) Eq. (12) 7: ˆSτ E[Sτ | C] Eq. (13) 8: Update θ via θ PTf τ=1 ψτ + ω Sτ ˆSτ 2 9: until converged Algorithm 2 Euler-Maruyama with Restart Inference Require: Solver step size 1/M 1: C Dtest, with C = [S Tp:0, C] 2: {µτ, στ, λτ, ντ, γτ} Tf τ=1 fθ(C) 3: t0 0, N M Tf 4: for i = 1, , N do 5: ti ti 1 + 1/M, ρti ti + 1 6: αi (µρti λρti kρti σ2 ρti /2)/M Drift 7: βi σρti z1/ M, with z1 N(0, 1) Diffusion 8: ζi κνρti + κγρti z2 with κ Pois(λρti /M), z2 N(0, 1) Jump 9: if (i 1) mod M = 0 then 10: ln Sti E[ln Sρti 1 | C] + αi + βi + ζi Restart 11: else 12: ln Sti ln Sti 1 + αi + βi + ζi 13: return { Sti}N i=1 |
| Open Source Code | Yes | Our code is available at https://github.com/DSL-Lab/neural-MJD. |
| Open Datasets | Yes | The S&P 500 dataset [89] is a public dataset containing historical daily prices for 500 major US stocks. |
| Dataset Splits | Yes | The data is split into 60% training, 20% validation, and 20% testing. |
| Hardware Specification | Yes | All experiments were run on NVIDIA A40 and A100 GPUs (48 GB and 80 GB VRAM, respectively). |
| Software Dependencies | No | The paper mentions various models and libraries like ARIMA [81], XGBoost [82], MLPs, GCNs [83], DDPM [45], EDM [84], FM [47], Neural CDE [33], Latent SDE [59], NJ-ODE [85], Autoformer [95], Ti DE [96], and N-Hi TS [97] through the Neural Forecast library (NIXTLA). However, it does not provide specific version numbers for any of these software components or frameworks used in the implementation. |
| Experiment Setup | Yes | We normalize the raw data into the range of [0, 1] for stability and use a regularization weight ω = 1.0 during training. All experiments were run on NVIDIA A40 and A100 GPUs (48 GB and 80 GB VRAM, respectively). ... For any time t Tf, we map it to an integer index via ρt := t + 1. ... we employ a sliding window approach with a stride of 1, using the past 14 days as input to predict the next 7 days. |