Nixtla Community #timegpt

Chathurangi Shyalika

08/29/2024, 2:41 AM

Hi All, Im trying to use detect_anomalies function as follows. anomalies_df = nixtla_client.detect_anomalies(df, time_col='time', target_col='M_Conv2_Speed_mmps', freq='D') It returns me below error. Can someone please help? Thank you.

Copy code

ApiError: status_code: 500, body: <!DOCTYPE html>
<!--[if lt IE 7]> <html class="no-js ie6 oldie" lang="en-US"> <![endif]-->
<!--[if IE 7]>    <html class="no-js ie7 oldie" lang="en-US"> <![endif]-->
<!--[if IE 8]>    <html class="no-js ie8 oldie" lang="en-US"> <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en-US"> <!--<![endif]-->
<head>
<title>Worker threw exception | api.nixtla.io | Cloudflare</title>
<meta charset="UTF-8" />
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8" />
<meta http-equiv="X-UA-Compatible" content="IE=Edge" />
<meta name="robots" content="noindex, nofollow" />
<meta name="viewport" content="width=device-width,initial-scale=1" />
<link rel="stylesheet" id="cf_styles-css" href="/cdn-cgi/styles/cf.errors.css" />
<!--[if lt IE 9]><link rel="stylesheet" id='cf_styles-ie-css' href="/cdn-cgi/styles/cf.errors.ie.css" /><![endif]-->
<style>body{margin:0;padding:0}</style>


<!--[if gte IE 10]><!-->
<script>
  if (!navigator.cookieEnabled) {
    window.addEventListener('DOMContentLoaded', function () {
      var cookieEl = document.getElementById('cookie-alert');
      cookieEl.style.display = 'block';
    })
  }
</script>
<!--<![endif]-->


</head>
<body>
  <div id="cf-wrapper">
    <div class="cf-alert cf-alert-error cf-cookie-error" id="cookie-alert" data-translate="enable_cookies">Please enable cookies.</div>
    <div id="cf-error-details" class="cf-error-details-wrapper">
      <div class="cf-wrapper cf-header cf-error-overview">
        <h1>
          <span class="cf-error-type" data-translate="error">Error</span>
          <span class="cf-error-code">1101</span>
          <small class="heading-ray-id">Ray ID: 8ba922e11cf7b46f • 2024-08-29 02:33:25 UTC</small>
        </h1>
        <h2 class="cf-subheadline" data-translate="error_desc">Worker threw exception</h2>
      </div><!-- /.header -->

      <section></section><!-- spacer -->

      <div class="cf-section cf-wrapper">
        <div class="cf-columns two">
          <div class="cf-column">
            <h2 data-translate="what_happened">What happened?</h2>
            <p>You've requested a page on a website (api.nixtla.io) that is on the <a href="<https://www.cloudflare.com/5xx-error-landing/>" target="_blank">Cloudflare</a> network. An unknown error occurred while rendering the page.</p>
          </div>

          
          <div class="cf-column">
            <h2 data-translate="what_can_i_do">What can I do?</h2>
            <p><strong>If you are the owner of this website:</strong><br />you should <a href="<https://www.cloudflare.com/login?utm_source=error_100x>" target="_blank">login to Cloudflare</a> and check the error logs for api.nixtla.io.</p>
          </div>
          
        </div>
      </div><!-- /.section -->

      <div class="cf-error-footer cf-wrapper w-240 lg:w-full py-10 sm:py-4 sm:px-8 mx-auto text-center sm:text-left border-solid border-0 border-t border-gray-300">
  <p class="text-13">
    <span class="cf-footer-item sm:block sm:mb-1">Cloudflare Ray ID: <strong class="font-semibold">8ba922e11cf7b46f</strong></span>
    <span class="cf-footer-separator sm:hidden">•</span>
    <span id="cf-footer-item-ip" class="cf-footer-item hidden sm:block sm:mb-1">
      Your IP:
      <button type="button" id="cf-footer-ip-reveal" class="cf-footer-ip-reveal-btn">Click to reveal</button>
      <span class="hidden" id="cf-footer-ip">34.66.126.43</span>
      <span class="cf-footer-separator sm:hidden">•</span>
    </span>
    <span class="cf-footer-item sm:block sm:mb-1"><span>Performance & security by</span> <a rel="noopener noreferrer" href="<https://www.cloudflare.com/5xx-error-landing>" id="brand_link" target="_blank">Cloudflare</a></span>
    
  </p>
  <script>(function(){function d(){var b=a.getElementById("cf-footer-item-ip"),c=a.getElementById("cf-footer-ip-reveal");b&&"classList"in b&&(b.classList.remove("hidden"),c.addEventListener("click",function(){c.classList.add("hidden");a.getElementById("cf-footer-ip").classList.remove("hidden")}))}var a=document;document.addEventListener&&a.addEventListener("DOMContentLoaded",d)})();</script>
</div><!-- /.error-footer -->


    </div><!-- /#cf-error-details -->
  </div><!-- /#cf-wrapper -->

  <script>
  window._cf_translation = {};
  
  
</script>

</body>
</html>

Chathurangi Shyalika

08/29/2024, 2:28 PM

Hi All, I am having an error in using forecast method. This timeseries dataset has rows of 2 mins apart. timegpt_fcst = nixtla_client.forecast(model='timegpt-1-long-horizon',finetune_steps=150, df=df, h=2000,freq='100ms', level=[90], time_col='formatted_date', target_col='x4', ) Error: ApiError: status_code: 500, body: {'status': 500, 'data': {'detail': 'Internal server error, Please contact us at ops@nixtla.io or Azul'}, 'message': 'error', 'details': 'request had an error', 'code': 'B10', 'support': 'If you have questions or need support, please email ops@nixtla.io', 'requestID': 'PNZFPRTHNV', 'headers': {}} Can someone please help? Thank you!

Santosh Puvvada

09/06/2024, 8:18 AM

Hi, I was using Fine tuning option in TimeGPT on a monthly data set. I was getting this error that each unique time series should have atleast 36 months data.. But in reality, very often we find time series with less than 3 years/36 months history. How to navigate through it? Any work around? Like, if i have 1000 time series in total & if 700 have >= 36 months, i can fine tune on them & then use the model for zero shot inferencing on other 300 time series or something like that?

Tracy Teal

09/04/2024, 8:06 PM

We have an exciting new update! 🚀 Introducing v2 of our TimeGPT API: Faster, Smarter, and More Powerful! 🚀 We’re thrilled to unveil the latest release of our API—v2, packed with incredible improvements driven by the community’s feedback. The nixtla API lets you seamlessly connect with TimeGPT. Just pip install the newest version of our SDK (v0.6.0), nixtla, and you will have access to these new features. With v2, we’ve focused on what matters most: speed, scalability, flexibility, and precision. Whether you’re working on anomaly detection, forecasting, or cross-validating TimeGPT, these enhancements will enable you to achieve better results, faster. ⚡ Unmatched Speed Improvements One of the standout upgrades in v2 is the dramatic increase in computational performance. We’ve fine-tuned our algorithms and optimized our infrastructure, delivering staggering results: we can detect anomalies 8.9x faster, forecast with exogenous variables 10x, and cross-validation 6x faster than the v1 of our API. These speedups aren’t just numbers—they represent a huge leap in efficiency, allowing you to run complex analyses in a fraction of the time. This is especially crucial in production environments where time-to-insight is key. ⏱️ 🌐 1 Billion Time Series in 6 Hours But that’s not all. With v2, we’ve shattered previous limits. In our latest experiment, we successfully forecasted 1 billion time series in just 6 hours. This unprecedented capability sets a new standard for scalability in time series forecasting, empowering organizations to handle massive datasets with unparalleled speed. 🚀 📊 Advanced Handling of Exogenous Variables We’ve also introduced a highly requested feature: the ability to distinguish between future and historical exogenous variables. Why is this important? - Historical Exogenous Variables: You can now leverage past data to boost the accuracy of your models, even when future data isn’t available. This is crucial for making reliable forecasts based on incomplete information. 🔍 - Future Exogenous Variables: When future data is available, you can fine-tune your forecasts even further, giving you a predictive edge. This dual approach enables more robust and adaptable models, helping you better anticipate trends and anomalies. 📈 🔍 Enhanced Model Explainability with SHAP Values In v2, we’ve also integrated SHAP values to enhance model interpretability. SHAP values allow you to understand the impact of each feature on TimeGPT’s predictions, providing deeper insights into the decision-making process. This is particularly valuable for model explainability and trust, especially in critical applications. 🧠 🛠️ New Integration with Polars In addition to these improvements, we’ve added support for *Polars*—a lightning-fast DataFrame library. With Polars, you can process large datasets more efficiently, making it easier to manage and manipulate your time series data. This perfectly complements our existing integrations with Dask, Ray, Spark, and Pandas. Why Polars? - Speed: Polars is built for performance, especially with large datasets. ⚡ - Memory Efficiency: It uses less memory, making it ideal for big data applications. 💾 - Parallelism: Polars automatically parallelizes operations, speeding up your data processing workflows. 🚄 What This Means for You We’ve listened to your feedback and made the changes you need to push the boundaries of what’s possible with time series forecasting. We’re eager to hear your thoughts and continue improving. 💙 Happy forecasting!

🎉 2

🚀 3

👍 2

Aravind Cheruvu

09/30/2024, 7:06 PM

I have a quick question, is there a way to apply TimeGPT to a series without having

ds

(timestamp) and with just index values 0,1,2,...?

Vidar Ingason

10/07/2024, 8:23 PM

I'm getting this again and again for various datasets in R.

Copy code

nixtla_client_fcst
[1] TimeGPT ds     
<0 rows> (or 0-length row.names)

I'm using nixtlar 0.5.4

Mariana Menchero

10/07/2024, 8:25 PM

Hi @Vidar Ingason today we're releasing the new version of nixtlar (v.0.6.0), which uses the new API for TimeGPT. In a few hours it should be in main.

Vidar Ingason

10/08/2024, 8:49 AM

I'm able to run TimeGPT again now. However, I was trying 4 hour data and got this error:

Copy code

timegpt_tbl
# A tibble: 13,032 × 3
   ds                  unique_id     y
   <dttm>              <chr>     <dbl>
 1 2022-04-14 03:00:00 id_333     34.5
 2 2022-04-14 07:00:00 id_333     37  
 3 2022-04-14 11:00:00 id_333     48  
 4 2022-04-14 15:00:00 id_333     41.4
 5 2022-04-14 19:00:00 id_333     40.1
 6 2022-04-14 23:00:00 id_333     28  
 7 2022-04-15 03:00:00 id_333     35.7
 8 2022-04-15 07:00:00 id_333     36  
 9 2022-04-15 11:00:00 id_333     42  
10 2022-04-15 15:00:00 id_333     41.7
# ℹ 13,022 more rows
# ℹ Use `print(n = ...)` to see more rows
> 
> fcst <- nixtla_client_forecast(df = timegpt_tbl, h = 12, id_col = "unique_id", time_col = "ds", target_col = "y")
Frequency chosen: 4h
Error in seq.Date(from = start_date, by = r_freq, length.out = h + 1) : 
  invalid string for 'by'
In addition: Warning message:
All formats failed to parse. No formats found.

Apoorv Khanuja

10/10/2024, 4:51 PM

Hello Nixtla team, I love the new dashboard look. However, I was wondering if there's a way to access daily/weekly API calls and especially the number of times API calls errored out? I have been getting "Invalid HTTP response: 504" errors and I'm trying to figure out if the API call also errored out or was it just my browser/the third party app I'm making the API calls from. Thank you!

Vidar Ingason

10/12/2024, 10:34 PM

Hi team, Is there any reason this is an error and not just a warning?

Copy code

Frequency chosen: M
Error in nixtlar::nixtla_client_forecast(df = timegpt_fit_tbl, h = prepared_data_list$horizon,  : 
  Your time series is too short. Please make sure that each of your series contains at least 36 observations.

Mariana Menchero

10/17/2024, 5:35 PM

We're very excited to share the latest updates to nixtlar, our R package for working with TimeGPT. If you're an R user, please check it out! Don’t forget to give us a ⭐ on GitHub if you find it useful! https://www.linkedin.com/feed/update/urn:li:activity:7252727877174009856/

https://www.youtube.com/watch?v=HO1qjhuPtfs▾

🙌 5

Vidar Ingason

10/18/2024, 1:06 PM

Hi team, I'm using the cross validation function and it seems that the cutoff point is incorrect. I have monthly data but the cutoff seems to be in seconds:

Copy code

> nixtla_client_cross_validation(
+     df = timegpt_tbl,
+     h  = 3,
+     n_windows = 5)
Frequency chosen: M
   unique_id         ds              cutoff       y TimeGPT
1      <tel:000122023-07-01|00012 2023-07-01> 1970-01-01 05:25:09 4691367 4763307
2      <tel:000122023-08-01|00012 2023-08-01> 1970-01-01 05:25:09 4632109 4660824
3      <tel:000122023-09-01|00012 2023-09-01> 1970-01-01 05:25:09 3925616 4002058
4      <tel:000122023-10-01|00012 2023-10-01> 1970-01-01 05:26:41 3282128 3235828
5      <tel:000122023-11-01|00012 2023-11-01> 1970-01-01 05:26:41 2579690 2739404
6      <tel:000122023-12-01|00012 2023-12-01> 1970-01-01 05:26:41 2446004 2606284

Jelte Bottema

10/21/2024, 11:40 AM

Hi Nixtla Team, I get very good results with the cross validation method, but I get not even comparable way worse results using the forecast method with time-gpt-1-long-horizon model even when I try to forecast the same time periods I forecasted with the cross validation. Is here a reason for? I am using exactly the same data setup for both methods, and in my understanding the results should be somewhat similar as it trains on the same previous data when trying to forecast the time periods and with the same model setup with the same finetunesteps and finetune loss function etc. So am I missing something? Like the difference is consistently way worse for foracast which does a very crappy forecast repeating the same "seasonality" forecast over and over for x number of days. And with cross validation it actually predicts a very good price prediction in the future which is not a repeated forecast of the same over and over again.

Apoorv Khanuja

10/21/2024, 9:32 PM

Hi Team, I was batch running timegpt forecast using 6 months of hourly data for training and 1 week of hourly data for testing for the last 2 years of data using sliding windows (2yr x 52weeks/yr = 104 API calls). Somewhere around the 30th API call, the function failed with the following error message:

status_code: 422, body: detail=[ValidationError(loc=['body', 1], msg='JSON decode error', type='json_invalid', ctx={'error': 'unexpected content after document'})]

. Could someone please shed some light on why I might be getting this error? Thanks for the help!

David McCandless

10/24/2024, 5:29 PM

Nixtla team, do y’all have suggested best practices on finetuning? like, iterate in steps of 10 b/w 0 and 100 e.g. 0, 10, 20, 30, … 100? Also, are you able to share any more color on what exactly

clean_ex_first=True

does with exogenous variables?

Paul Reichert

10/30/2024, 11:22 AM

Can't log in into dashboard.nixtla.io/sign_in I tried different emails. Received the confirmation email, but I was not logged in when I opened it.

Paul Reichert

11/03/2024, 4:33 PM

Hi all, anyone with a tip on how to make the finetune depth work?

Jon T

12/06/2024, 4:44 PM

Hi there, I've got some promising good results using TimeGPT. So good, I feel I should be skeptical 🤨 😂 It appears the main caveats are (and I hope there are inaccuracies/workarounds/solutions here ;): 1. Fine-tuning can't be carried forward between runs, thus if we have a data stream every second or minute (for many series), we'd have to run the entire pipeline every time. 2. This inherently introduces latency, and our primary optimization path would be to reduce size/complexity of data 3. Does using AzureAI and/or Spark assist? I suppose each series can be run in parallel, but does Azure enable project-specific fine-tuning/training? Impressive product, though indeed. Hoping I can find a way to make it work

Stuart Kerr

12/16/2024, 7:15 AM

FYI: Quick singular use case using 2-years worth of data (1-year input and 1-year test). Compared a highly tweaked (and trained) Neural-Prophet model versus out of the box, no tweaks one-shot TimeGPT; the MAPE difference was only 1.4% higher using TimeGPT. Great job Nixtla.

🙌 1

👍 2

Micah Denver

12/17/2024, 6:34 PM

Hi, I was wondering why TimeGPT doesn’t allow missing data/timestamps? Is that something that will be added in the future? Thanks

Jon T

12/22/2024, 8:18 PM

The docs mention the historical forecast horizon != horizon; instead, fixed based on freq. Is there a mapping somewhere to see what each freq is mapped to? This would be extremely helpful in establishing benchmarks. Thx!

José Morales

01/02/2025, 9:18 PM

Hey folks. We just released version 0.6.5 of the nixtla package, which includes methods for saving and reusing finetuned models. Please upgrade the package (

pip install -U nixtla

), give it a try and let us know if you run into any issues. • Tutorial • Full release notes

👍 4

🙌 2

Seethamraju Purvaj

01/03/2025, 12:35 PM

Hi Team, I am trying to forecast some financial data. I see that I get the same accuracy / predictions even if I add the holiday columns and date columns or remove them. Just wanted to know if my hunch that the nixtla sdk internally uses these dates (holidays) by default ? or is that my feature selection is not affecting the values. Also I see that I can't use the

feature_contributions

as when I try to extract them I get nixtla client has no parameter / member variable feature_contributions do we need any specific version of nixtla for this ? I have

0.6.4

Thanks.

Jerry Gamblin

01/09/2025, 5:45 PM

Hi Team! I am using Nixtla to make some CVE growth predictions and would like to use the graphs, but there doesn’t seem to be a documented way to change the plot title from

unique_id=ts_0

. Am I missing something obvious?

Alexandre Moreau

01/13/2025, 4:40 PM

Hey everyone! Using TimeGPT in Azure, I get this error message when trying to finetune the model: 'Minimum number of samples by id required for finetuning is 25, got 24.' I validated that I have enough datapoints (176) and that the number of datapoints is okay for what the documentation says for finetuning. With monthly data, we minimally need 48 datapoints + h + step_size +(or *) (n_windows - 1). with h = 4, step_size=1 and n_windows=20, I need 72 datapoints? That's not what the error says + I have more than enough datapoints. I am kinda lost, is there something I am missing? Thank you!

Mike Powers

01/22/2025, 6:07 PM

Hey Nixtla Team - We've been experiencing some difficulty with adding historical exogenous variables using hist_exog_list from the documentation. It seems the function call is not recognizing the variable.

Mike Powers

01/22/2025, 6:07 PM

forecastraw = nixtla_client.forecast(df=modeldf, h=monthspredict, finetune_steps=200, finetune_loss='mae', id_col='unique_id', time_col='Month', target_col='Activity',hist_exog_list=['Exogenous1', 'Exogenous2']) * TypeError: NixtlaClient.forecast() got an unexpected keyword argument 'hist_exog_list'

Mike Powers

01/22/2025, 6:09 PM

Previously we had just been adding the historical exogenous variables to our dataframe modeldf and thought the model was using them automatically until we noticed our models with and without historical exogenous variables were giving us the same forecasts. Were there any changes to the model recently? Is there another way to implement historical exogenous variables?

saaid

02/01/2025, 8:13 PM

Has anyone dealt with this error before? while using timeGPT

Copy code

[SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed: unable to get local issuer certificate (_ssl.c:992)

Sarah Unterseher

04/09/2025, 2:51 PM

Hi everyone, I have a question regarding the training data requirements for TimeGPT. With the following line of code: timegpt_fcst_df = nixtla_client.forecast(df=df, h=96, freq='15min', id_col='unique_id', time_col='ds', target_col='y') I get the following error:

Copy code

ValueError: Series contain missing or duplicate timestamps, or the timestamps do not match the provided frequency.
Please make sure that all series have a single observation from the first to the last timestamp and that the provided frequency matches the timestamps'.
You can refer to <https://docs.nixtla.io/docs/tutorials-missing_values> for an end to end example.

My training data (df) is structured as follows: I have three columns unique_id, ds and y. The time series of one unique_id are always 288 rows long and the next time series starts one hour after the previous one. So there are the same timestamps several times but with different unique_ids. According to the documentation, this should work, right? Unfortunately I don't understand why I am getting this error in my setup.