Tempo

Status: in development. Tempo is a new Numerical Method product that has not been released yet. This page describes what it is being built to do. There is no download, price or release date yet; to follow its progress or to discuss early access, contact us.

The problem

Machine-learning trading models are usually tuned over a grid of learning rates, epochs and penalties. As shown in Training Is Numerical Integration, much of that grid is redundant: for a wide class of models the learning rate and the number of steps act together through a single quantity, the training time $t = kh$, which behaves like a ridge penalty. Searching the redundant grid wastes compute and, worse, multiplies the number of configurations tried on the same data, which is exactly how backtests become optimistic.

The approach

Tempo makes training time the primary control and measures what it does:

• it tracks $t = kh$ and the stability margin $h\lambda_{\max}$ during training;
• it reports the effective degrees of freedom of the model along the training path, so that a model’s complexity can be stated as a number of effective parameters rather than as a list of hyperparameters;
• it chooses the stopping time out of time, on data that follows the training window, and records how many configurations were tried so that reported performance can be adjusted for multiple testing.

What it is being built to do

• Training-time and stability diagnostics for gradient-based training.
• Effective-degrees-of-freedom estimates for linear and kernel models, and local estimates for neural networks.
• Out-of-time early stopping and rolling re-validation.
• Regime-aware checks that flag when the validated stopping time drifts with market volatility.

How it fits

Tempo is intended to sit next to AlgoQuant’s strategy optimization and backtesting tools, bringing the same discipline to machine-learning models that the strategy optimization article argued for in classical strategies.