I'm new to the platform so apologies if my syntax is incorrect - here to learn
I've been trying to make forecasts with some monthly data that seems to be heavily impacted by COVID. The models DataRobot are giving me seems to perform worse than a traditional TS approach (I used Excel for this) and I just wanted to know if there is a way DataRobot can provide traditional TS models only (I tried the "manual" model selection but I could not find any models other than the baseline model that resembles a traditional TS model.
I have tried tweaking the best performing models ranked according to the error metric too but I think there is insufficient data post COVID to make any reliable predictions - I stand to be corrected though.
Any help on this issue would be really appreciated 🙂
Solved! Go to Solution.
Cannot get into the provided link Linda:
Hi @Linda, it certainly is a great reply.
I have actually seen this article and a few others, as well as some videos after digging around in the community forums. I think the issue was that it was a lot of information to take in initially and, because I'm still fairly new to DataRobot, I was not sure if I was missing something when things went wrong. But @matt's response really helped deconstruct what I can do to improve the model and troubleshoot problems that may occur.
Thanks for that really really informative reply, I have a way better understanding of the models DataRobot implements on the data - I will look into the model selection process further and play around with it to gain an even better understanding too.
Regarding the adjustments that can be made to help improve the chosen model, everything you mentioned in the last paragraph I initially took into consideration (varying FDW, including a calendar, making some features KIA, varying forecast period). However, I overlooked splitting up the project into two, which in hindsight I should've done as well, as they do have variable behaviour - so thank you for highlighting that point. Hopefully that does the trick in terms of improving the overall accuracy of the model.
Again, thanks so much for this great reply!
Looks like great information from @matt ! 🥳 You may want to have a look at this community article, Automated Time Series Walkthrough, if you haven't found it already.
Also here's a quick image to show where to configure backtesting - fyi
The baseline models are mainly designed to give you a model that uses the last known data point to predict the next datapoint. You may hear this referred to as a "naive" approach when talking to different people. There are several different options for doing a "traditional" time series approach that you could try from the repository on your data, as well as a few other options which might help.
They may not be obvious as traditional time series initial but if you look in the repository you will see models that use have ARIMA, AUTOARIMA, Mean Response Regressor, and other names in them. You can also use the icons to help differentiate these. I can explain the differences in some of the major model types which might help also.
Some other things you might consider looking at which could help with training are using calendar features to indicate the start of Covid in your modeling data, or possibly removing some historical data using the backtests or trimming the data manually before loading it. This can help to prevent the model learning to much about the past if it is no longer relevant to what is happening now. In the Advanced Options at the start screen you can configure the number and length of your backtests. You can use this to make sure that your training and validation periods are relevant and that you have good coverage over the covid time window, and that it is no entirely excluded from training (e.g. all covid is in your holdout or validation + holdout) with none in training.
One final thing you may consider doing is placing series that behave very differently from one another into two separate projects. I'm not 100% sure what type of data you have, but for sales data, if you have products with very low or intermittent sales, and products with a fairly regular sales cycle, breaking the training data up and creating two projects can also lead to a lot of accuracy boost.