Add tide forecasting example using NNS.ARMA

[OVVO-Financial]

Summary

This PR adds a complete gist demonstrating tide forecasting using the ovvo-nns library's NNS.ARMA algorithm. The example reproduces Lord Kelvin's 1872 tide prediction machine results but discovers seasonal patterns empirically from raw water-level data rather than relying on hand-tuned astronomical constituents.

Changes

• tides_nns.py: Main forecasting script that:

  • Loads NOAA tide-gauge data from a remote CSV
  • Splits data into training (pre-2019-07-15) and test (2-week holdout) sets
  • Uses nns_seas() to detect seasonal periods from the training data
  • Filters periods to only those usable by the optimizer (< training_set / 4)
  • Runs nns_arma_optim() to fit an ARMA model with seasonal factors
  • Evaluates forecast accuracy (R² ≈ 0.965) and generates visualization
  • Includes timing instrumentation for performance tracking

• README.md: Documentation covering:

  • Historical context of Lord Kelvin's tide machine (1872)
  • Overview of the NNS.ARMA approach vs. physics-based methods
  • Results summary with selected periods and R² metric
  • Installation and execution instructions
  • Explanation of the non-obvious period filtering logic

• tides_forecast.png: Generated visualization showing actual vs. predicted tides with 95% prediction intervals

Notable Implementation Details

• Period filtering is applied before slicing to top-100 candidates, ensuring the optimizer receives usable periods rather than oversized ones that would be rejected
• Uses lin_only=True for linear-only optimization
• Validation window equals the forecast horizon (2 weeks) for realistic out-of-sample evaluation
• Runtime dominated by optimizer (~2 min); seasonality detection is near-instant

https://claude.ai/code/session_01MxVLKYqazC2uuAW3P3MbAm