Forecasting Time Series Data at Scale, with the TICK stack Nathaniel Cook @nathanielvcook 

Overview ● Time series data forecasting ● Challenges of scale ● Introduce Facebook's Prophet procedure ● Example using the TICK stack 

What is time series data forecasting? ● Predict future values based off past values ● Compute accuracy using historical windows ● Use simple baseline models 

What are the challenges of scale? Not enough time and resources to manage each individual series 

What is Facebook's Prophet procedure? ● Algorithm ● Workflow 

What is the Prophet algorithm? Simple general additive model: y(t) = g(t) + s(t) + h(t) ● g(t) - growth ● s(t) - seasonality ● h(t) - holidays Simple intuitive parameters for each term 

What is the Prophet workflow? Create/Update Model Evaluate Model Analyst in the Loop Surface Problems Visually Inspect Model 

Example ● Dataset ● Goals 

Dataset ● Github stars for ~400 python projects ● Projects came from an awesome-python list ● Small and large projects ● Diverse uses of python from common libraries to end user applications ● New and old projects 

Goals ● Forecast all 400 time series reliably ● Repair any problematic forecasts ● Do it in a scalable manner 

What is the TICK stack? ● Telegraf - collection agent (not used in today's example) ● InfluxDB - database ● Chronograf - visualization ● Kapacitor - processing engine 

How does the TICK stack enable the workflow? ● Store time series in InfluxDB. ● Use Kapacitor tasks to evaluate and surface problems with models. ● Use Chronograf to inspect models. 

All Github Stars by Project 

1. Create/Update Model ● 3 Kapacitor task templates ○ Baseline models: Mean, Exponential Smoothing ○ Prophet Model (UDF) ○ 3 tasks per Github project = ~1200 tasks ● Each task(project) can have its own parameters 

Model Prophet Task Template var data = batch |query(''' SELECT value FROM srcDB.srcRP.srcMeasurement WHERE project = '$project' ''') .period(history) .every(forecast) .align() .groupBy(groupBy) @prophet() .periods(forecast / interval) .field('value') .changepointPriorScale(changepointPriorScale) .intervalWidth(uncertaintyIntervalWidth) |influxDBOut() 

2. Evaluate Models ● Evaluate each Github project for the past ~10 years of data, for each model type (mean, holt, prophet) ● Compute accuracy of each project/model using "mean absolute percentage error" (MAPE) 

Accuracy Task var errors = src |join(forecasted) .on('project') |eval(lambda: abs(("src.value" - "forecasted.value") / "src.value")).as('error') var sum_errors = errors |sum('error').as('value') var count = errors |count('error').as('value') sum_errors |join(count) .as('sum_errors', 'count') |eval(lambda: float("sum_errors.value") / float("count.value")) .as('mape') |influxDBOut() 

3. Surface Problematic Models // Best Performers SELECT bottom(mape, project, model, 10) FROM star_counts WHERE time > now() - 30d AND model = 'prophet' // Worst Performers SELECT top(mape, project, model, 10) FROM star_counts WHERE time > now() - 30d AND model = 'prophet' 

How did the first pass go? 

4. Visually Inspect Models 

1. Update the Model 

How does it look now? 

Summary ● Forecasting at scale is about reducing the cost per forecast ● Using simple models and automating the workflow enables forecasting at scale ● The TICK stack provides a platform on which to automate the workflow 

Resources ● https://docs.influxdata.com/ ● https://github.com/vinta/awesome-python ● https://facebookincubator.github.io/prophet/ ● https://www.gnu.org/software/parallel/ Questions? 

Extras Live Explorations http://localhost:3000