SuperAPI AI Directory

timesfm-forecaster

General & other

View repository

Updated 10-16 06:18

About this skill

Google Research TimesFM Forecaster Skill

Role Overview

Provides a predictive price projection layer using Google Research's state-of-the-art TimesFM model (decoder-only Transformer-based, trained on billions of time series data points). Replaces subjective guesswork with measurable mathematical statistical intervals.

Quantitative Capabilities

  1. Zero-Shot Forecasting:
    • Accepts the last 128 - 512 candles (OHLCV) and projects 16 - 32 candle horizons ahead without needing to retrain the model (fine-tuning from scratch).
  2. Dynamic Trailing Stop via Quantile P10:
    • Calculates the pessimistic lower-bound estimate (Quantile 10th percentile).
    • Used as the benchmark for a Dynamic Trailing Stop Loss that moves up following the statistical normality deviation.
  3. Realistic Profit Targets via P50 & P90:
    • The median line (P50) serves as the most rational price target estimate.
    • The optimistic line (P90) serves as a partial take-profit target to secure peak profits.
  4. Hardware-Adaptive Inference:
    • Runs on NVIDIA GPU acceleration (PyTorch CUDA) when available, or automatically falls back to a memory-efficient CPU mode.