BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//pyconde-pydata-2026//talk//B8KVNJ
BEGIN:VTIMEZONE
TZID:CET
BEGIN:STANDARD
DTSTART:20001029T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20000326T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-pyconde-pydata-2026-B8KVNJ@pretalx.com
DTSTART;TZID=CET:20260415T105500
DTEND;TZID=CET:20260415T112500
DESCRIPTION:Forecasting talks love a clean ending: “and then we improved 
 WMAPE by 3.7%.”\nNice. Now put that model into production without suffer
 ing from instability.\n\nYou retrain your model on a few new weeks of data
  and suddenly the one-year forecast jumps 15–20%. Planning teams redo de
 cisions\, trust erodes\, and your “accurate” model becomes unusable. T
 his talk is about forecast stability: how much forecasts change when you a
 dd new data and rerun the same pipeline.\n\nWe run a simple experiment: tr
 ain a model\, forecast one year ahead\, add recent data\, retrain\, and me
 asure forecast-to-forecast change. We repeat this across common forecastin
 g approaches including ETS/ARIMA\, Prophet\, XGBoost with lag features\, A
 utoGluon ensembles\, neural/global models\, and TimeGPT-style APIs.\n\nYou
  will see that high accuracy does not guarantee usable forecasts\, and tha
 t some models are systematically more volatile than others. We then cover 
 practical ways to stabilise forecasts without freezing them\, focusing on 
 reconciliation and ensembling (including origin ensembling).\n\nThis talk 
 is for forecasting practitioners who want models users actually trust\, no
 t just good metrics.
DTSTAMP:20260523T181858Z
LOCATION:Helium [3rd Floor]
SUMMARY:Accuracy Is Overrated: Ship Stable Forecasts (Without Lying to Your
 self) - Illia Babounikau
URL:https://pretalx.com/pyconde-pydata-2026/talk/B8KVNJ/
END:VEVENT
END:VCALENDAR