Join us on Thursday for a deep dive into the Nixtla Engine:
Marco will run a live forecast from Claude using Nixtla MCP and show the auto-researcher, a team of agents that searches for the best model for your data and goals.
Save your spot: nixtla.io/webinar-events/ins…
"Inside the Nixtla Engine: Agentic Forecasting, Live"
Thursday, November 19
8 AM PT · 11 AM ET · 5 PM CET
Sign up below ⬇️
Can't make it? Register and we’ll send you the recording.
nixtla.io/webinar-events/ins…
See you next week?
Your forecasting pipeline still needs you. It just doesn’t need you to try 40 configurations by hand.
On Thursday, Marco Peixeiro will show how this works in the Nixtla Engine. He sets the evaluation and the boundaries. The Engine runs the experiments and reports back with the pipeline that scored best, ready for production from the first run. You check what drove each forecast and ship it.
Bring your questions, Marco will answer them live.
🗓️ Thursday, November 19
⏱️ 8:00 AM PT / 11:00 AM ET / 17:00 CET
🔗 Register here: nixtla.io/webinar-events/ins…
Can’t join live? Register anyway to access the recording later.
If there’s something you’d like Marco to cover, reply below and we’ll bring it into the session.
See you there!
A production forecast is never just a model call. Getting the pipeline right takes repetitive trial and error, and that’s exactly the kind of work agents can take over.
Data scientists clean inputs, engineer features, pick models, apply business rules, ensemble, reconcile across hierarchies, and evaluate the options to find the best setup.
In our next webinar, Applied AI Scientist Marco Peixeiro will show how the Nixtla Engine lets you describe those steps in a few lines, instead of coding and wiring them up yourself.
Since the Engine can run and evaluate the pipelines you describe, an agent can do the same trial and error a data scientist would: swap models, test features, try reconcilers, and keep what works.
Marco will cover:
- how to ask the Engine questions through your own LLM
- a preview of the auto-researcher, which runs experiments to search for the best pipeline
- the Engine’s latest explainability capabilities, which help you understand not only what a forecast predicts but why
7 years ago I started my "financial career" by forecasting the S&P 500 with LSTM RNNs with tensorflow.
LSTMs were still fairly new back then.
I spent a lot of time thinking about each of the hyperparameters (activation function, number of layers and neurons, training epochs, dropout rate, optimizer).
Recently I wanted to see how far the field has come. So I tried @nixtlainc , which my friends Max and Cristian have been building for the past 5 years. And also backed by @Microsoft !!!
One of their most advanced models is TimeGPT-2, a transformer-based foundation model trained on over 100 billion data points. You point it at a time series, pass a few parameters, and you get a forecast. The API even does embedded fine-tuning.
Crazy how far I can get with a couple lines of code, what took me weeks in the past ....
So, obviously I applied TimeGPT-2 to a few macro data series and the experience was seamless. It feels like at some point we might have Forecasting as a Service lmao
Anyway, I made the entire experiment open source, and it runs as an @openbb_finance App that you can add directly to the OpenBB Workspace. The recipe is also available so it's easier to extend to different timeseries.
Let me know what you think!
🎉 Announcing Nixtla Enterprise 2.0 🎉
Tl;DR: more models, domain expertise, reasoning capabilities, mcp interactions and optimized compute environments.
We’re excited to share the next iteration of our enterprise offering. Starting today, companies can sign up for early access. 🚀
These new features are designed to help engineers build, evaluate, and deploy forecasting pipelines more effectively.
🌐 In Nixtla Enterprise 2.0, time series models, LLMs, agents, and humans work together through three core capabilities:
1️⃣ Model Zoo: built with the best models
- Optimized implementations of leading foundation models, including Chronos 2 (Amazon Web Services (AWS)), TimesFM 2.5 (Google), and FlowState-r1 (IBM)
- Enterprise-ready, battle-tested models from the Nixtlaverse, spanning statistical, ML, and neural approaches
2️⃣ Unified Interface: simple, consistent integration
- Add or swap models in your pipelines by changing a single line of code
3️⃣ Time Series Agent: AI-assisted forecasting, redefined
- Trained on years of experience deploying state-of-the-art forecasting systems at leading companies
- Plan and run end-to-end pipelines in your favorite IDE or AI provider
- Powered by Nixtla MCP, which provides domain knowledge and a fully integrated execution environment so agents can reason, generate code, run analysis, refine results, and make recommendations
Together, these capabilities unlock a new way to build, compare, and operationalize time series intelligence:
🧠 Guided generation and iterative refinement informed by domain expertise and the latest research.
⚙️ Adaptive support across models and strategies based on performance feedback.
🔄 End-to-end experimentation workflows that reduce manual overhead while keeping humans in control.
🤝 Flexibility to use natural language or Python and plug Nixtla into existing AI workflows.
This launch marks a new chapter in Nixtla’s mission: building a time-series ecosystem that blends adaptive tooling with human expertise, helping teams forecast and iterate more efficiently with the latest innovations.
📩 Join the waitlist to be among the first to try this version
👇 Links to demos and blog post
#HappyForecasting#timeseries#forecasting#AI#LLM
Training a custom model for every cryptocurrency you want to forecast is tedious and impossible to scale.
There are thousands of tokens. Market conditions shift constantly. By the time your model is trained, the opportunity is gone.
TimeGPT changes this with zero-shot forecasting: predict new cryptocurrencies without any training on them.
Researchers tested this on 21 cryptocurrencies (BTC, ETH, SOL, and 18 others):
• Lowest average error across all 21 assets
• MAPE of 2.7% on daily data, 0.7% on hourly data
• No training required per cryptocurrency
• Outperformed TFT, TiDE, and PatchTST statistically
This research by Wang, Braslavski, and Ignatov demonstrates how foundation models are changing financial forecasting.
🚀 Full paper: bit.ly/4orAfGX#TimeGPT#TimeSeries#Cryptocurrency
The same anomaly detection model can flag 89 or 505 anomalies, depending on one parameter.
At 99% confidence, TimeGPT only flags extreme outliers.
Drop it to 70%, and you catch subtle shifts that might indicate early warning signs.
Neither is "correct." It depends on whether you want fewer false positives or fewer missed detections.
This article by Khuyen Tran covers the end-to-end process to:
• Set up TimeGPT for anomaly detection
• Run detection on real Wikipedia traffic data
• Add exogenous features to catch context-dependent patterns
• Tune sensitivity from 99% to 70% confidence
🚀 Full article: bit.ly/44L48uS#TimeGPT#AnomalyDetection#TimeSeries#Nixtla
When you’re forecasting many time series, each one has its own pattern.
A single model won’t capture all that complexity, so you shouldn’t rely on one model across the entire dataset.
But testing several models per series by hand is tedious and impossible to scale.
StatsForecast handles this automatically by training multiple statistical models in parallel and using cross-validation to choose the best one for each time series.
This article walks through the full workflow:
• How to define several automatic models
• How rolling cross-validation works
• How StatsForecast picks the best model per series
• How to generate final forecasts with uncertainty
🚀 Full article: bit.ly/49FMY5o#timeseries#forecasting#python#statsforecast
How do you tell if recent behavior is normal variation or a fundamental business shift?
MLForecast's Combine class lets you compose any two transformations with operators to create features like:
• Baseline deviation (subtraction): Track drift from historical norms
• Momentum ratios (division): Detect acceleration and deceleration
• Volatility measures (std/mean): Identify unstable periods
• Cumulative signals (addition): Aggregate multiple time windows
In the example below, we create a baseline deviation feature by subtracting the expanding average from the 7-day rolling mean. The deviation signal shows you:
⬆️ Positive deviation: Recent values consistently higher than historical average
⬇️ Negative deviation: Recent values consistently lower than historical average
➡️ Zero deviation: Operating at historical norm
🚀 Full tutorial on lag transformation: bit.ly/4o4g5me#MLForecast#TimeSeries#FeatureEngineering#DataScience
Automate date feature engineering with TimeGPT 🚀
For time series with seasonal patterns like holidays, weekday effects, or monthly trends, date-based features can significantly improve forecast accuracy.
However, manually creating these features is time-consuming and requires writing code to extract and encode date components.
TimeGPT's date_features parameter automatically generates temporal features from timestamps and one-hot encodes them.
The plot below shows the forecast with date features (blue) tracks actual values (white) more accurately than without date features (pink).
🚀 Full tutorial: bit.ly/4p6J0a6#FeatureEngineering#TimeSeriesAnalysis#DataScience#TimeGPT
Quantify forecast uncertainty with MLForecast conformal prediction intervals 📊
Point forecasts provide no measure of prediction uncertainty.
Without confidence intervals, analysts cannot assess reliability or make probability-based decisions.
MLForecast provides conformal prediction intervals that quantify forecast uncertainty with statistical guarantees.
How to generate conformal prediction intervals:
• Pass PredictionIntervals(n_windows, h) to fit() for calibration
• Set level parameter in predict() to specify confidence levels
🚀 Full guide: bit.ly/4oSijGp#DataScience#MachineLearning#TimeSeries#Python
Detect production anomalies in minutes with TimeGPT's rolling forecast 🚨
Traditional anomaly detection requires retraining on entire datasets, making real-time monitoring impossible.
This leads to delayed responses when critical system failures or data quality issues occur in production.
TimeGPT solves this with online anomaly detection that analyzes recent data windows in real-time using rolling forecasts.
How to use:
• Call detect_anomalies_online() with your data
• Set detection_size (how many recent data points to use)
• Configure h (forecast steps) and level (confidence)
🚀 Full guide: bit.ly/4oGsxJB#TimeSeries#AnomalyDetection#TimeGPT#AI
Automate NeuralForecast hyperparameter tuning with Ray Tune ⚡️
Manual hyperparameter tuning for neural forecasting models can be time-consuming and requires deep expertise in model architectures.
NeuralForecast's Auto models (AutoNHITS, AutoTFT, AutoLSTM) automatically search hyperparameter spaces using Ray Tune or Optuna.
Key capabilities:
• Automated hyperparameter optimization
• Multiple search backends (Ray Tune, Optuna)
• Built-in neural architectures (NHITS, TFT, LSTM)
• Parallel search with Ray Tune
🚀 Full guide: bit.ly/4pONALq#DataScience#AutoML#TimeSeries#Forecasting