BiTCN — Bidirectional Temporal Convolutional Network

Fast training · Bidirectional TCN · NeuralForecast-native

---

Overview

• Model ID: bitcn-forecasting
• Architecture: BiTCN (NeuralForecast / NIXTLA)
• Framework: NeuralForecast (PyTorch Lightning backend)
• Backend: forecasting specialist (specialists-ts-cu121)
• specialist_type: forecasting
• Hardware: GPU (small footprint, ~512 MB VRAM); also runs on CPU (slower)
• Training time: 2–20 minutes (fast convergence)

BiTCN processes a series in both forward and backward directions with dilated causal convolutions, giving a large receptive field without the quadratic cost of attention — a strong, fast baseline.

---

When to Use

✅ Perfect for:

• Fast training iteration
• Univariate forecasting (single series)
• Multi-series forecasting with the NeuralForecast unique_id format
• A strong baseline without Transformer complexity

❌ Not ideal for:

• Complex multivariate dependencies — try PatchTST
• Zero-shot forecasting — use TimesFM 2.5
• Very irregular timestamps — use Prophet

---

Training

Dataset format

BiTCN uses the NeuralForecast standard format — a CSV with three columns:

Column	Type	Description
unique_id	string	Series identifier (use "series_0" for a single series)
ds	datetime	Timestamp
y	float	Target value

unique_id,ds,y
series_0,2020-01-01,102.3
series_0,2020-01-02,98.7
store_A,2020-01-01,1523.0
store_B,2020-01-01,892.0

Launch training

from colabhive import ColabHive

client = ColabHive(api_key="YOUR_KEY", account_id="YOUR_ACCOUNT")

dataset = client.datasets.upload(name="my-sales-dataset", file="./train.csv")

job = client.training.create(
    model="bitcn-forecasting",
    dataset_id=dataset.id,
    hyperparameters={
        "h": 30,            # forecast horizon
        "input_size": 180,  # context window
        "frequency": "D",   # D=daily
        "learning_rate": 1e-3,
        "max_steps": 1000,
        "batch_size": 32,
        "dropout": 0.1,
    },
)
job.wait()
print(job.get_metrics())

# Publish the trained model as an inference endpoint
endpoint = client.training.register_for_inference(
    run_id=job.id,
    name="bitcn-sales-endpoint",
    description="BiTCN sales forecaster",
    visibility="account",
)

The exact accepted hyperparameters come from the template's schema: GET /api/builder/v1/training/model-configs/{model_config_id}/schema.

---

Inference

result = client.endpoints.infer(
    endpoint_id=endpoint.endpoint_id,
    input_data={
        "values": [102.3, 98.7, 105.1, 99.4, 103.8],  # historical values
        "horizon": 96,
        "frequency": "D",
    },
)
print(result)

curl -X POST "https://api.colabhive.com/api/builder/v1/endpoints/{ENDPOINT_ID}/infer" \
  -H "X-Account-ID: YOUR_ACCOUNT_ID" \
  -H "X-API-Key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input": {"values": [102.3, 98.7, 105.1, 99.4], "horizon": 30, "frequency": "D"}}'

---

Hyperparameter Guide

Parameter	Effect	Typical Range
h	Forecast horizon	1–720
input_size	Context window	2×h to 10×h
max_steps	Training iterations	500–5000
batch_size	Mini-batch size	16–256
dropout	Regularization	0.0–0.3

Frequency reference: H hourly · D daily · W weekly · M monthly · Q quarterly · T minute.

---

Related

• PatchTST Forecasting — Transformer alternative
• TimesFM 2.5 — zero-shot, no training needed
• Classical Forecasting — Prophet, ARIMA for simple patterns