Learning the Efficient Frontier
Authors: Philippe Chatigny, Ivan Sergienko, Ryan Ferguson, Jordan Weir, Maxime Bergeron
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We used a MC sampling scheme to cover the entire domain in table. 1 uniformly to generate a dataset of approximately one billion samples Dtrain = [(Zinput,1, EF((Zinput,1), ], which we used to train Neural EF in a supervised fashion: ... We generated two test datasets Dtest, Dvalidation of 1 million samples each on the the same domain as the training set described in table 1. |
| Researcher Affiliation | Industry | Philippe Chatigny Riskfuel Toronto pc@riskfuel.com Ivan Sergienko Beacon Platform New York ivan.sergienko@beacon.io Ryan Ferguson Riskfuel Toronto rf@riskfuel.com Jordan Weir Riskfuel Toronto jw@riskfuel.com Maxime Bergeron Riskfuel Toronto mb@riskfuel.com |
| Pseudocode | Yes | The complete optimal allocation of eq. 3 can be summarized by the following python script: |
| Open Source Code | No | The paper does not provide any concrete access information (specific repository link, explicit code release statement, or code in supplementary materials) for the Neural EF model's source code. |
| Open Datasets | No | We used a MC sampling scheme to cover the entire domain in table. 1 uniformly to generate a dataset of approximately one billion samples Dtrain = [(Zinput,1, EF((Zinput,1), ], which we used to train Neural EF in a supervised fashion: ... and All synthetic datasets mimic real-life distributions.... No access information for this generated dataset is provided. |
| Dataset Splits | Yes | We generated two test datasets Dtest, Dvalidation of 1 million samples each on the the same domain as the training set described in table 1. |
| Hardware Specification | Yes | trained on a single NVIDIA A100 GPU with stochastic gradient descent using the Adam W optimizer [34] and the L2 loss. We also used an annealing learning rate decay starting from 5.5e 5 to 1.0e 6. As stated in sec. 2, we also implemented a baseline EF optimization in Py Torch that we used solely for comparing the evaluation throughput (evaluations/seconds) between Neural EF over the base pricer which was implemented using CVXOPT [14]. The hyperparameters of Neural EF are described in table. 2 and were selected by estimated guesses from the accuracy measured on Dvalidation. |
| Software Dependencies | No | implemented using CVXOPT [14], Pytorch. Specific version numbers for these software dependencies are not provided. |
| Experiment Setup | Yes | The hyperparameters of Neural EF are described in table. 2 and were selected by estimated guesses from the accuracy measured on Dvalidation. Table 2: Hyper Parameters of Neural EF. Token dimension 320 Transformer depth 8 Transformer # heads 8 Feed forward projection 1024 Output activation Sigmoid Embedding method [30, 24] Hidden activation Swish [42] and also with stochastic gradient descent using the Adam W optimizer [34] and the L2 loss. We also used an annealing learning rate decay starting from 5.5e 5 to 1.0e 6. |