Forecasting

Soltix provides built-in time-series forecasting with 7 algorithms (6 statistical + 1 auto-selector) in Go, plus optional ML models via the soltix-ml Python service.

Built-in Forecasters (Go)

Algorithm	Type	Best For
sma	Simple Moving Average	Stable, flat data
exponential	Exponential Smoothing	Data with noise
linear	Linear Regression	Data with clear trend
holt_winters	Triple Exponential Smoothing	Seasonal + trend data
arima	ARIMA(p,d,q)	Stationary or differenced data
prophet	Prophet-style decomposition	Complex seasonality (daily/weekly/yearly)
auto	Auto-selector	General use

Auto-Selection Logic

Condition	Selected Algorithm
Has seasonality + enough data	Holt-Winters
Has trend	Linear
≥20 data points	Exponential
Small dataset	SMA

API

Endpoints

Method	Path
GET	/v1/databases/:db/collections/:col/forecast
POST	/v1/databases/:db/collections/:col/forecast

Parameters

Parameter	Default	Description
device_id	—	Device IDs (required)
start_time	—	RFC3339 — historical data start (required)
end_time	—	RFC3339 — historical data end (required)
field	—	Single field to forecast
fields	—	Multiple fields (POST body)
algorithm	auto	sma, exponential, holt_winters, linear, arima, prophet, auto, ml
horizon	24	Number of periods to forecast
seasonal_period	24	For Holt-Winters
data_interval	1h	Data interval: 1m, 5m, 15m, 30m, 1h, 4h, 1d, 1w
ml_algorithm	—	When algorithm=ml: random_forest or lstm

Example

# Auto forecast for next 24 hours
curl "http://localhost:5555/v1/databases/mydb/collections/sensors/forecast?\
device_id=sensor-001&\
field=temperature&\
start_time=2026-01-01T00:00:00Z&\
end_time=2026-01-15T00:00:00Z&\
horizon=24&\
algorithm=auto" \
  -H "X-API-Key: your-api-key"

Response

{
  "results": [
    {
      "device_id": "sensor-001",
      "field": "temperature",
      "predictions": [
        {
          "timestamp": "2026-01-15T01:00:00Z",
          "value": 25.8,
          "lower_bound": 24.2,
          "upper_bound": 27.4
        }
      ],
      "model_info": {
        "algorithm": "holt_winters",
        "mape": 3.2,
        "mae": 0.8,
        "rmse": 1.1,
        "data_points": 336
      }
    }
  ]
}

ML Models (soltix-ml)

For heavier-weight predictions, the optional soltix-ml Python service trains machine learning models:

Model	Algorithm	Description
Random Forest	scikit-learn RF	Ensemble method, warm start, sliding window features
LSTM	PyTorch LSTM	Neural network with normalization baked into ONNX graph

Training Pipeline

1. Discovery: Connect to Router → list databases, collections, devices
2. Data: Load 30 days of CSV data per device
3. Features: Sliding window of configurable size
4. Training: Both RF and LSTM for every device × field
5. Export: ONNX format (input: (1, window_size) float32, output: (1, 1) float32)
6. Upload: Multipart form upload to Router

Using ML Forecasts

curl "http://localhost:5555/v1/databases/mydb/collections/sensors/forecast?\
device_id=sensor-001&\
field=temperature&\
algorithm=ml&\
ml_algorithm=lstm&\
horizon=24" \
  -H "X-API-Key: your-api-key"