eulogik/nanoforecast-v03

🔮 NanoForecast v0.3
_{World's most deployable time series transformer}

6.5M params · 512 context · Streaming RNN · ONNX-ready
Runs on CPU, Raspberry Pi, and in the browser

---

🚀 Why NanoForecast?

Most time series foundation models are too big to ship — they need GPUs, terabytes of training data, and a PhD to deploy. NanoForecast is different:

Feature	NanoForecast	Other Foundation Models
Parameters	200K – 6.5M	200M – 10B+
Inference device	CPU, Raspberry Pi, browser	GPU required
Model size	1.4 MB (ONNX)	1 GB+
Streaming	✅ Stateful RNN — feed one value at a time	❌ Fixed-window only
Training	2 min on laptop	1000+ GPU-hours
ONNX export	✅ Built-in	❌ Often broken
Zero-shot	✅ on 6 benchmark datasets	✅ but expensive
License	Apache 2.0	Often restrictive

Stop renting GPUs for forecasting. Train on your laptop. Deploy to a $35 Raspberry Pi. Get production forecasts in minutes.

🏆 Benchmarks

Dataset	MASE	sMAPE (%)	MAE	CRPS
ETTh1	1.95	12.06	1.30	1.05
ETTh2	2.74	10.47	2.46	1.99
ETTm1	2.17	10.70	0.72	0.65
exchange_rate	7.44	1.72	0.011	0.014
electricity	1.29	4.76	158.30	175.24
traffic	0.81	24.00	0.004	0.003
Overall	2.73	10.62	27.13	29.83

Benchmarks on 6 standard datasets. Overall MASE improved 21% from v0.2 (3.45 → 2.73). See GitHub for full results on 3 model sizes.

📦 Quick Start

pip install nanoforecast

import numpy as np
from nanoforecast import NanoForecast

model = NanoForecast.from_pretrained("eulogik/nanoforecast-v03")
context = np.sin(np.linspace(0, 8*np.pi, 512)) + 0.1 * np.random.randn(512)
out = model.predict(context, horizon=48, freq=1)
print(out["forecast"].shape)  # (48,) point forecast

🔄 Streaming Inference (Unique to NanoForecast)

NanoForecast's DeltaNet RNN maintains a recurrent state across calls — no other TS model does this.

result = model.predict(context, horizon=48, return_state=True)
state = result.pop("state")

for new_val in incoming_data_stream:
    result = model.predict_step(new_val, state, horizon=48)
    print(result["forecast"][0, :5])  # updated forecast instantly

Use it for:

• Real-time IoT sensor monitoring
• Live financial tick data
• Interactive dashboards
• Edge devices with limited memory

🎯 Try It in 1 Click

Upload a CSV → get forecast + prediction intervals + decomposition. No code. No GPU.

🧠 Model Details

Attribute	Value
Profile	d96-L8
Parameters	6,518,104
Context length	512
Prediction length	48
Hidden dim / layers	96 / 8
Architecture	LongConv + DeltaNet RNN + Gated Router + MLP
Outputs	Point forecast + p10/p25/p50/p75/p90 quantiles
Decomposition	Trend + Seasonal + Residual
Export	ONNX FP32 + INT8
Streaming	Stateful DeltaNet RNN

⚡ Training Details

Attribute	Value
Datasets	ETTh1, ETTh2, ETTm1, exchange_rate, electricity, traffic
Synthetic records	10,000
Epochs	200
Best epoch	147
Val loss (best)	0.2230
Batch size	128
Learning rate	3e-5
Wall time	42,144s (11.7h on Colab T4)

✨ What Makes NanoForecast Special

• 🪶 Featherweight: 200K–6.5M params, not 200M+. Ships as a dependency, not an API call.
• 📡 Streaming native: Feed one value at a time, get updated forecasts. Perfect for IoT and real-time data.
• 💻 Zero-GPU inference: Runs on ARM, x86, RISC-V. Benchmarked at <50ms on Raspberry Pi 4.
• 📦 Single pip install: pip install nanoforecast → ready in 30 seconds.
• 🏭 Production-grade: ONNX export, Docker, FastAPI server all included.
• 📊 Full uncertainty: 5 quantiles + prediction intervals + trend/seasonal decomposition.
• 🔌 HuggingFace native: from_pretrained / push_to_hub — works like any HF model.

🔧 Deploy Anywhere

# FastAPI server
pip install nanoforecast fastapi uvicorn python-multipart
python3 deploy/fastapi_server.py

# ONNX export (edge/IoT/browser)
pip install "nanoforecast[onnx]"
python3 -m nanoforecast.export.onnx_export \
    --checkpoint <dir> \
    --output nanoforecast.onnx

# Docker
docker build -t nanoforecast .
docker run -p 8000:8000 nanoforecast

📚 Pretrained Variants

Model	Params	Size	Context	Best For
nanoforecast-200k	~676K	2.7 MB	256	Extreme edge / RPi Zero
nanoforecast-500k	~1.6M	6.4 MB	256	General purpose / mobile
nanoforecast-v03 (you are here)	~6.5M	26 MB	512	Max accuracy / server

📄 Paper

Read the tech report: deploy/paper.tex (LaTeX, 8 sections).

⚠️ Known Limitations

• Modest accuracy vs. 200M+ param models (MASE 2.73 vs TimesFM ~0.72)
• Fixed context window (512 for v0.3)
• Univariate channels (no cross-channel attention)
• Training requires PyTorch (inference can be pure ONNX)

❤️ Built by Eulogik

Eulogik builds deployable AI for the real world. We believe ML models should run on the hardware people actually have, not the hardware vendors want to sell.

• 🌐 eulogik.com
• 🐙 GitHub
• 📦 PyPI

Star the repo ⭐ if you find this useful — it helps others discover NanoForecast!