TimesFM Forecasting — Setup Guide¶
Source: k-dense-ai/scientific-agent-skills (699 installs) Also available from: google-research/timesfm (153), eturkes/claude-scientific-skills (23) Category: Data Science / Forecasting
Google Research's TimesFM is a decoder-only foundation model for time series forecasting. Unlike traditional statistical models (ARIMA, Prophet), TimesFM requires zero training — pass historical data and get predictions immediately. The skill wraps this into a Claude Code / Hermes agent workflow.
Installation¶
npx skills add k-dense-ai/scientific-agent-skills --skill timesfm-forecasting
For the Google Research variant:
npx skills add google-research/timesfm --skill timesfm-forecasting
Prerequisites¶
| Requirement | Details |
|---|---|
| Python 3.10+ | python3 --version — TimesFM requires 3.10+ |
| Historical data | CSV or JSON with timestamp + value columns |
| Hermes Agent | Any version with skills support |
Install TimesFM Python package:
pip install timesfm
The skill will install this automatically on first use, but pre-installing avoids cold-start latency.
Key Capabilities¶
Core Features¶
| Capability | How to Trigger | Notes |
|---|---|---|
| Point forecasting | "Forecast next 30 days of this revenue data" | Zero-shot — no model training |
| Uncertainty intervals | "Forecast with 90% confidence intervals" | Built-in quantile prediction |
| Multi-horizon | "Predict 7-day, 30-day, and 90-day forecasts" | Variable horizon from single call |
| Seasonal decomposition | "Decompose this time series" | Trend + seasonal + residual |
| Anomaly detection | "Find anomalies in this usage data" | Flags points outside prediction intervals |
CLI Command Reference¶
# Point forecast
python3 SKILL_DIR/scripts/forecast.py --input data.csv --horizon 30
# With confidence intervals
python3 SKILL_DIR/scripts/forecast.py --input data.csv --horizon 90 --ci 0.95
# Anomaly detection
python3 SKILL_DIR/scripts/forecast.py --input data.csv --detect-anomalies
CorpusIQ Use Cases¶
| Use Case | How |
|---|---|
| Revenue forecasting | Predict monthly revenue from Stripe/QuickBooks export data |
| Usage growth prediction | Forecast agent session growth, API call volume |
| Churn risk detection | Identify accounts with declining usage patterns |
| Capacity planning | Predict compute/storage needs from historical metrics |
| ROI projection | Model expected returns from growth campaigns |
Troubleshooting¶
| Issue | Fix |
|---|---|
ModuleNotFoundError: timesfm |
pip install timesfm — package not auto-installed |
CUDA not available |
TimesFM runs on CPU. Add --device cpu flag for non-GPU environments. |
| Out-of-memory on large datasets | TimesFM context window is 512 points. Batch longer series or downsample. |
| Prediction looks flat | Input series may lack clear trend/seasonality. Add exogenous features or extend history. |
Verification¶
# Verify skill installed
hermes skills list | grep timesfm-forecasting
# Quick test with sample data
python3 -c "
import pandas as pd
import numpy as np
dates = pd.date_range('2026-01-01', periods=100, freq='D')
values = np.sin(np.arange(100) * 0.3) + np.random.normal(0, 0.1, 100) + np.arange(100) * 0.05
pd.DataFrame({'ds': dates, 'y': values}).to_csv('/tmp/test_series.csv', index=False)
print('Sample data written to /tmp/test_series.csv')
"
← Skills Catalog | June 30 Update 2 Discovery → Powered by CorpusIQ