Editing Openai/693eaf5a-1494-8007-b1e3-00abefb11523 (section)

=== The most promising approach is not a single fixed geometric formula, but a parameterized tetrahedral family whose free parameters are fit to historical data via walk-forward backtesting. Concretely: ===
* Parameterize the tetrahedral apex height and the projection/scale factors (these control how much time vs price contribute).
* Fit those parameters by minimizing an error metric (MSE or MAPE) on an in-sample set with walk-forward validation.
* Use rolling windows (e.g., last 90 days) to form tetrahedra and predict the next day; aggregate predictions (ensemble) to stabilize noisy results.

Below are the concrete model families, optimization method, backtest plan, and evaluation metrics.