Time Series Forecasting and Nixtla Ecosystem!

Python Milano meetup

Milan, May 18th 2023

github/pietroppeter

👨‍👩‍👧 Me - Pietro Peterlongo

past: Math (Pisa, Trieste), Climate Science (Paris), Cryptography (Trento)

now (since 2015): Principal Data Scientist at ToolsGroup: Supply Chain Planning Software

Python: started early 2000s, serious around 2015, my first presentation in... May 2023 ;)

...also active in Open Source with Nim

Why Forecasting?

What is Time Series Forecasting?

Domains of applications

events planning (how many?)

weather forecasting (sun or rain?)

economics (growth next year?)

control theory (reactor will heat?)

finance (stock up or down?)

web analytics (how much traffic?)

energy (how much consumption?)

supply chain (how many sales?)

How to forecast

equations with $t$ (ODEs, PDEs, ...)

judgmental forecasts

statistical forecasting

classical (ETS, ARIMA, ...)

ml (random forest, LGBM, ...)

neural (NBEATS, DeepAR, ...)

ensemble forecasting

Dealing with Uncertainity

Supply Chain

A Crash Course

Methodology

1. think about your why
2. gather data (process, explore)
3. baseline
4. measure
5. improve
6. restart from step 4 or less

Baseline

• Historical average
• Naive (aka persistence in weather forecasting)
• Moving Average
• Seasonal Naive
• ..., existing forecast (benchmark)

Metrics

1. low
2. median
3. average

from Forecast KPIs: RMSE, MAE, MAPE & Bias

$$ \text{MAPE} = \frac{1}{N} \Sigma \frac{\left| e_t \right|}{d_t} $$

$$ \text{MAE} = \frac{1}{N} \Sigma \left| e_t \right| $$

$$ \text{RMSE} = \sqrt{\frac{1}{N} \Sigma e_t^2} $$

Cross Validation (aka Backtesting)

ETS (Error, Trend, Seasonal)

AutoETS performs automatic model selection

References

FPP: Forecasting Principles and Practices

by Rob J Hyndman and George Athanasopoulos

Monash University, Australia

3rd edition, May 2021

free! otexts.com/fpp3/

R-based

Other references

• Youtube (+ Kaggle) course by Konrad Banchewicz (Feb 2022): video 1, kaggle 0
• Time Series Analytics and Forecasting, Lectures by Lorenzo Nespoli (May 2023)
• Demand Forecasting con Serie Temporali gerarchiche, intervista Pythonista Ep10 (Nov 2021)

Python Libraries for Timeseries

Hyndman's suggestions

The best Python implementations for my time series methods are available from Nixtla.

• statsmodels
• sktime
• skforecast
• Darts
• Prophet (from Facebook)
• GluonTS (from Amazon)
• Merlion (from Salesforce)
• ..., awesome-time-series

Nixtla Story

Hi! At nixtla, we have just launched our MVP: Open Source Forecasting Pipeline. We want to democratize access to state-of-the-art forecasting models. Feedback and comments are welcome! https://t.co/1y7ttPYNgy

— fede garza (they/them) (@fede_gr) October 15, 2021

I'm thrilled to announce that we're releasing the fastest #autoarima implementation for #Python today! 😍

It is a mirror implementation of @robjhyndman's #autoarima (R) and it is optimized using @numba_jit. It is 20x faster than pmdarima and @MetaAI's Prophet. 😎

🧵@nixtlainc

— fede garza (they/them) (@fede_gr) February 22, 2022

Forecast 1M time series in 30 minutes. @nixtlainc + @raydistributed + @numba_jit = 🤩 🚀 ⚡

Recently, we launched the fastest version of #AutoARIMA for #python. Now you can scale your computation horizontally for millions of time series, leveraging the power of @raydistributed pic.twitter.com/zkfBL0FYJ4

— fede garza (they/them) (@fede_gr) May 3, 2022

🎉Today at @nixtlainc, we are excited to release our new 👑HierarchicalForecast library. 🔥

You can simply reconcile forecasts for hierarchical problems using statistical approaches such as Bottom Up, Top Down, Middle Out, Minimum Trace, and ERM. #Python #Forecasting

— fede garza (they/them) (@fede_gr) July 26, 2022

🎉 We are thrilled to announce the release of the latest version of mlforecast a #Python library for Scalable #machinelearning 🤖 for #timeseries #forecasting

🚀 This version comes with exciting new features that are sure to make forecasting even more efficient and accurate

🧵 pic.twitter.com/UR7wM0DTUJ

— nixtla (@nixtlainc) February 21, 2023

Team

Why Nixtla?

• open source
• api
• performance
• vision

Risks: moving fast, iterating on business model

Nixtlaverse

import lightgbm as lgbm
from mlforecast import MLForecast
from window_ops.expanding import expanding_mean
from window_ops.rolling import rolling_mean

mlf = MLForecast(
    models = [lgbm.LGBMRegressor()], # list of models
    freq = 'MS', # frequency of time series: month start
    differences=[12], # differences to apply to target
    lags=[1, 12], # lags to use as feature
    lag_transforms=( # lag transformation to apply to specific lags
        1: [expanding_mean],
        12: [(rolling_mean, 24)],
    )
)

A declarative feature engineering pipeline!

mlf.fit(df)
forecast_df = mlf.predict(12, levels=[90])