NanoForecast v0.5: From MASE 2.73 → 1.326 with Zero Architecture Changes

How fixing training data pipelines and loss computation slashed error by 51% on a 6.5M parameter model.

The Problem

Time series forecasting models fall into two camps:

1. Massive foundation models (TimesFM, Chronos, Lag-Llama) — require GPUs, 100M+ parameters, and can't run on edge devices
2. Small deployable models — tiny and fast, but accuracy is usually mediocre

NanoForecast v0.5 proves you can have both: 6.5M parameters, CPU inference, ONNX export, streaming — AND MASE 1.326 (competitive with models 10× larger).

The Key Insight

We didn't change the architecture between v0.3 and v0.5. Same LongConv + DeltaNet RNN, same gated router, same MLP blocks. The 51% improvement came entirely from fixing the training pipeline:

1. Corrected loss computation — v0.3 had a subtle bug where the reconstruction loss was computed on mismatched tensor shapes
2. Fixed tensor truncation — multi-horizon training was truncating context tensors that should have been preserved
3. Better data mixing — 6 real datasets + 10K synthetic records, properly shuffled with resolution-aware batching
4. Proper training recipe — OneCycleLR with 3e-5 peak LR, gradient clipping at 1.0, bfloat16 mixed precision

This is a lesson for the community: sometimes the biggest gains come from fixing your data pipeline, not your model.

Benchmark Results

The model achieves MASE < 1.0 on 3 of 6 datasets — a threshold that previously required models 10-100× larger.

Architecture: Why It Works

NanoForecast uses a hybrid architecture that combines the best of transformers and RNNs:

• LongConv captures global periodic patterns (weekly, monthly cycles)
• DeltaNet RNN captures local dependencies and maintains streaming state
• Gated Router dynamically blends the two representations per layer
• Monotonic Quantile Head guarantees p10 ≤ p25 ≤ p50 ≤ p75 ≤ p90 (physically meaningful uncertainty)

The conservation identity ensures: trend + seasonal + residual ≡ point forecast (exact, not approximate).

Deployment: From Laptop to Raspberry Pi

# Install
pip install nanoforecast

# Inference (anywhere)
import numpy as np
from nanoforecast import NanoForecast

model = NanoForecast.from_pretrained("eulogik/nanoforecast-v05")
result = model.predict(context, horizon=48, freq=1)

Deployment options:

Streaming inference (unique to NanoForecast)

# Stream new observations one at a time
result = model.predict(context, horizon=48, return_state=True)
state = result.pop("state")

for new_val in incoming_stream:
    result = model.predict_step(new_val, state, horizon=48)
    # Updated forecast instantly

No other time series model supports this. The DeltaNet RNN maintains a recurrent state that encodes all past observations.

Training on Your Own Data

# From CSV in 2 minutes
python3 train_from_csv.py --csv sales.csv --target revenue --horizon 48

# Or full pretraining on multiple datasets
python3 pretrain.py \
  --datasets ETTh1,ETTh2,exchange_rate,electricity \
  --epochs 200 \
  --batch-size 128 \
  --device cuda \
  --output checkpoints/nanoforecast-v05

Ablation toggles in the Colab notebook:

• USE_DART_NORM = False/True — DART-Norm (causal mean/std normalization)
• USE_MULTI_HORIZON = False/True — Random horizon lengths during training

Comparison with Other Models

What's Next

• v0.6: DART-Norm integration (causal normalization) + multi-horizon training
• v0.7: Multivariate cross-series dependencies
• v0.8: OpenRouter API — $0.001/forecast
• v1.0: Production-ready with monitoring and drift detection

Resources

• Model: eulogik/nanoforecast-v05
• GitHub: github.com/eulogik/NanoForecast
• Live Demo: huggingface.co/spaces/eulogik/nanoforecast
• Colab Training: Open in Colab

Citation

@article{nanoforecast2026,
  title={NanoForecast v0.5: A Deployable Time Series Foundation Model},
  author={Eulogik},
  year={2026},
  url={https://github.com/eulogik/NanoForecast}
}

Built by Eulogik — deployable AI for the real world. If you found this useful, please ⭐ the GitHub repo and like the model on Hugging Face!