Prophet

When to use Prophet

Prophet works best when your data meets most of the following conditions:

• At least two full seasonal cycles of historical data (for example, two years of daily observations)

• Clear seasonal patterns—daily, weekly, or yearly

• Known one-off events that affect the series (product launches, holidays, promotions)

• A growth trend approaching a natural ceiling or floor (for example, market saturation)

If your data has none of these characteristics, a simpler model may perform just as well.

How it works

1. Load time series data into an MTable column using the MTable Assembler component.

2. Pass the MTable column to the Prophet component and configure forecasting parameters.

3. Prophet decomposes the series into trend, seasonality, and holiday components, then generates a forecast.

4. The output contains a point forecast (yhat) in the prediction column and uncertainty bounds (yhat_lower, yhat_upper) in the prediction details column.

Configure the component in the PAI console

Configure the component by using code

Copy the following code into the PyAlink Script component to replicate Prophet component behavior programmatically.

The code performs three steps:

1. Aggregate raw time series rows into a single MTable column per group using GroupByBatchOp.

2. Run Prophet forecasting on each MTable using ProphetBatchOp.

3. Flatten the MTable prediction output into a regular table using FlattenMTableBatchOp.

import time, datetime
import numpy as np
import pandas as pd

# Download the required Prophet plugin environment
downloader = AlinkGlobalConfiguration.getPluginDownloader()
downloader.downloadPlugin('tf115_python_env_linux')

# Sample input: each row is one timestamped observation
data = pd.DataFrame([
    [1, datetime.datetime.fromtimestamp(1),  10.0],
    [1, datetime.datetime.fromtimestamp(2),  11.0],
    [1, datetime.datetime.fromtimestamp(3),  12.0],
    [1, datetime.datetime.fromtimestamp(4),  13.0],
    [1, datetime.datetime.fromtimestamp(5),  14.0],
    [1, datetime.datetime.fromtimestamp(6),  15.0],
    [1, datetime.datetime.fromtimestamp(7),  16.0],
    [1, datetime.datetime.fromtimestamp(8),  17.0],
    [1, datetime.datetime.fromtimestamp(9),  18.0],
    [1, datetime.datetime.fromtimestamp(10), 19.0]
])

# Step 1: Convert rows into an MTable column grouped by series ID
source = dataframeToOperator(data, schemaStr='id int, ts timestamp, val double', op_type='batch')

source.link(
    GroupByBatchOp()
        .setGroupByPredicate("id")
        .setSelectClause("id, mtable_agg(ts, val) as data")
    # Step 2: Run Prophet and forecast 4 future periods
).link(
    ProphetBatchOp()
        .setValueCol("data")
        .setPredictNum(4)
        .setPredictionCol("pred")
    # Step 3: Flatten the MTable prediction output into rows
).link(
    FlattenMTableBatchOp()
        .setSelectedCol("pred_detail")
        .setSchemaStr("ds timestamp, yhat double")
).print()

What's next

• Use the MTable Expander component to flatten predictionDetailCol output into a standard table for downstream analysis.

• For an overview of available components, see Overview of Machine Learning Designer and Component reference: overview of all components.