Benchmarking Lag-Llama towards XGBoost
22 hours in the past
On Hugging Face, there are 20 fashions tagged “time collection” on the time of writing. Whereas definitely not loads (the “text-generation-inference” tag yields 125,950 outcomes), time collection forecasting with basis fashions is an fascinating sufficient area of interest for giant corporations like Amazon, IBM and Salesforce to have developed their very own fashions: Chronos, TinyTimeMixer and Moirai, respectively. On the time of writing, one of the crucial in style on Hugging Face by variety of likes is Lag-Llama, a univariate probabilistic mannequin. Developed by Kashif Rasul, Arjun Ashok and co-authors [1], Lag-Llama was open sourced in February 2024. The authors of the mannequin declare “sturdy zero-shot generalization capabilities” on a wide range of datasets throughout completely different domains. As soon as fine-tuned for particular duties, additionally they declare it to be the very best general-purpose mannequin of its sort. Massive phrases!
On this weblog, I showcase my expertise fine-tuning Lag-Llama, and take a look at its capabilities towards a extra classical machine studying method. Particularly, I benchmark it towards an XGBoost mannequin designed to deal with univariate time collection knowledge. Gradient boosting algorithms akin to XGBoost are extensively thought-about the epitome of “classical” machine studying (versus deep-learning), and have been proven to carry out extraordinarily properly with tabular knowledge [2]. Subsequently, it appears becoming to make use of XGBoost to check if Lag-Llama lives as much as its guarantees. Will the inspiration mannequin do higher? Spoiler alert: it’s not that easy.
By the way in which, I cannot go into the small print of the mannequin structure, however the paper is value a learn, as is that this good walk-through by Marco Peixeiro.
The info that I take advantage of for this train is a 4-year-long collection of hourly wave heights off the coast of Ribadesella, a city within the Spanish area of Asturias. The collection is obtainable on the Spanish ports authority data portal. The measurements have been taken at a station positioned within the coordinates (43.5, -5.083), from 18/06/2020 00:00 to 18/06/2024 23:00 [3]. I’ve determined to mixture the collection to a every day degree, taking the max over the 24 observations in every day. The reason being that the ideas that we undergo on this publish are higher illustrated from a barely much less granular perspective. In any other case, the outcomes turn out to be very risky in a short time. Subsequently, our goal variable is the utmost peak of the waves recorded in a day, measured in meters.
There are a number of the explanation why I selected this collection: the primary one is that the Lag-Llama mannequin was skilled on some weather-related knowledge, though not loads, comparatively. I might anticipate the mannequin to search out such a knowledge barely difficult, however nonetheless manageable. The second is that, whereas meteorological forecasts are usually produced utilizing numerical climate fashions, statistical fashions can nonetheless complement these forecasts, specifically for long-range predictions. On the very least, within the period of local weather change, I believe statistical fashions can inform us what we’d usually anticipate, and the way far off it’s from what is definitely taking place.
The dataset is fairly customary and doesn’t require a lot preprocessing aside from imputing just a few lacking values. The plot beneath exhibits what it appears to be like like after we break up it into practice, validation and take a look at units. The final two units have a size of 5 months. To know extra about how we preprocess the info, take a look at this notebook.
We’re going to benchmark Lag-Llama towards XGBoost on two univariate forecasting duties: level forecasting and probabilistic forecasting. The 2 duties complement one another: level forecasting offers us a selected, single-number prediction, whereas probabilistic forecasting offers us a confidence area round it. One may say that Lag-Llama was solely skilled for the latter, so we must always deal with that one. Whereas that’s true, I consider that people discover it simpler to grasp a single quantity than a confidence interval, so I believe the purpose forecast continues to be helpful, even when only for illustrative functions.
There are lots of components that we have to take into account when producing a forecast. A number of the most vital embrace the forecast horizon, the final commentary(s) that we feed the mannequin, or how usually we replace the mannequin (if in any respect). Completely different mixtures of things yield their very own kinds of forecast with their very own interpretations. In our case, we’re going to do a recursive multi-step forecast with out updating the mannequin, with a step measurement of seven days. Which means we’re going to use one single mannequin to supply batches of seven forecasts at a time. After producing one batch, the mannequin sees 7 extra knowledge factors, similar to the dates that it simply predicted, and it produces 7 extra forecasts. The mannequin, nonetheless, isn’t retrained as new knowledge is obtainable. When it comes to our dataset, which means we are going to produce a forecast of most wave heights for every day of the following week.
For level forecasting, we’re going to use the Mean Absolute Error (MAE) as efficiency metric. Within the case of probabilistic forecasting, we are going to purpose for empirical protection or coverage probability of 80%.
The scene is ready. Let’s get our fingers soiled with the experiments!
Whereas initially not designed for time collection forecasting, gradient boosting algorithms usually, and XGBoost particularly, might be nice predictors. We simply have to feed the algorithm the info in the fitting format. As an illustration, if we need to use three lags of our goal collection, we will merely create three columns (say, in a pandas dataframe) with the lagged values and voilà! An XGBoost forecaster. Nevertheless, this course of can rapidly turn out to be onerous, particularly if we intend to make use of many lags. Fortunately for us, the library Skforecast [4] can do that. The truth is, Skforecast is the one-stop store for growing and testing all types of forecasters. I actually can’t suggest it sufficient!
Making a forecaster with Skforecast is fairly simple. We simply have to create a ForecasterAutoreg object with an XGBoost regressor, which we will then fine-tune. On prime of the XGBoost hyperparamters that we’d usually optimise for, we additionally have to seek for the very best variety of lags to incorporate in our mannequin. To try this, Skforecast gives a Bayesian optimisation methodology that runs Optuna on the background, bayesian_search_forecaster.
The search yields an optimised XGBoost forecaster which, amongst different hyperparameters, makes use of 21 lags of the goal variable, i.e. 21 days of most wave heights to foretell the following:
Lags: [ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21]
Parameters: {'n_estimators': 900,
'max_depth': 12,
'learning_rate': 0.30394338985367425,
'reg_alpha': 0.5,
'reg_lambda': 0.0,
'subsample': 1.0,
'colsample_bytree': 0.2}
However is the mannequin any good? Let’s discover out!
Level forecasting
First, let’s have a look at how properly the XGBoost forecaster does at predicting the following 7 days of most wave heights. The chart beneath plots the predictions towards the precise values of our take a look at set. We will see that the prediction tends to comply with the overall development of the particular knowledge, however it’s removed from excellent.
To create the predictions depicted above, we’ve used Skforecast’s backtesting_forecaster operate, which permits us to judge the mannequin on a take a look at set, as proven within the following code snippet. On prime of the predictions, we additionally get a efficiency metric, which in our case is the MAE.
Our mannequin’s MAE is 0.64. Which means, on common, our predictions are 64cm off the precise measurement. To place this worth in context, the usual deviation of the goal variable is 0.86. Subsequently, our mannequin’s common error is about 0.74 models of the usual deviation. Moreover, if we have been to easily use the earlier equal commentary as a dummy greatest guess for our forecast, we’d get a MAE of 0.84 (see level 1 of this notebook). All issues thought-about, evidently, up to now, our mannequin is healthier than a easy logical rule, which is a reduction!
Probabilistic forecasting
Skforecast permits us to calculate distribution intervals the place the longer term end result is prone to fall. The library gives two strategies: utilizing both bootstrapped residuals or quantile regression. The outcomes are usually not very completely different, so I’m going to focus right here on the bootstrapped residuals methodology. You may see extra ends in half 3 of this notebook.
The thought of establishing prediction intervals utilizing bootstrapped residuals is that we will randomly take a mannequin’s forecast errors (residuals) an add them to the identical mannequin’s forecasts. By repeating the method various occasions, we will assemble an equal variety of different forecasts. These predictions comply with a distribution that we will get prediction intervals from. In different phrases, if we assume that the forecast errors are random and identically distributed in time, including these errors creates a universe of equally attainable forecasts. On this universe, we’d anticipate to see at the very least a proportion of the particular values of the forecasted collection. In our case, we are going to purpose for 80% of the values (that’s, a protection of 80%).
To assemble the prediction intervals with Skforecast, we comply with a 3-step course of: first, we generate forecasts for our validation set; second, we compute the residuals from these forecasts and retailer them in our forecaster class; third, we get the probabilistic forecasts for our take a look at set. The second and third steps are illustrated within the snippet beneath (the primary one corresponds to the code snippet within the earlier part). Strains 14-17 are the parameters that govern our bootstrap calculation.
The ensuing prediction intervals are depicted within the chart beneath.
An 84.67% of values within the take a look at set fall inside our prediction intervals, which is simply above our goal of 80%. Whereas this isn’t dangerous, it might additionally imply that we’re overshooting and our intervals are too huge. Consider it this manner: if we stated that tomorrow’s waves could be between 0 and infinity meters excessive, we’d at all times be proper, however the forecast could be ineffective! To get a thought of how huge our intervals are, Skforecast’s docs counsel that we compute the world of our intervals by thaking the sum of the variations between the higher and decrease boundaries of the intervals. This isn’t an absolute measure, however it will possibly assist us evaluate throughout forecasters. In our case, the world is 348.28.
These are our XGBoost outcomes. How about Lag-Llama?
The authors of Lag-Llama present a demo notebook to start out forecasting with the mannequin with out fine-tuning it. The code is able to produce probabilistic forecasts given a set horizon, or prediction size, and a context size, or the quantity of earlier knowledge factors to think about within the forecast. We simply have to name the get_llama_predictions operate beneath:
The core of the funtion is a LagLlamaEstimatorclass (traces 19–47), which is a Pytorch Lightning Estimator primarily based on the GluonTS [5] package deal for probabilistic forecasting. I counsel you undergo the GluonTS docs to get acquainted with the package deal.
We will leverage the get_llama_predictions operate to supply recursive multistep forecasts. We merely want to supply batches of predictions over consecutive batches. That is what we do within the operate beneath, recursive_forecast:
In traces 37 to 39 of the code snippet above, we extract the percentiles 10 and 90 to supply an 80% probabilistic forecast (90–10), in addition to the median of the probabilistic prediction to get a degree forecast. If it is advisable study extra concerning the output of the mannequin, I counsel you take a look on the creator’s tutorial talked about above.
The authors of the mannequin advise that completely different datasets and forecasting duties might require differen context lenghts. In our case, we strive context lenghts of 32, 64 and 128 tokens (lags). The chart beneath exhibits the outcomes of the 64-token mannequin.
Level forecasting
As we stated above, Lag-Llama isn’t meant to calculate level forecasts, however we will get one by taking the median of the probabilistic interval that it returns. One other potential level forecast could be the imply, though it might be topic to outliers within the interval. In any case, for our specific dataset, each choices yield comparable outcomes.
The MAE of the 32-token mannequin was 0.75. That of the 64-token mannequin was 0.77, whereas the MAE of the 128-token mannequin was 0.77 as properly. These are all increased than the XGBoost forecaster’s, which went right down to 0.64. The truth is, they’re very near the baseline, dummy mannequin that used the earlier week’s worth as in the present day’s forecast (MAE 0.84).
Probabilistic forecasting
With a predicted interval protection of 68.67% and an interval space of 280.05, the 32-token forecast doesn’t carry out as much as our required customary. The 64-token one, reaches an 74.0% protection, which will get nearer to the 80% area that we’re in search of. To take action, it takes an interval space of 343.74. The 128-token mannequin overshoots however is nearer to the mark, with an 84.67% protection and an space of 399.25. We will grasp an fascinating development right here: extra protection implies a bigger interval space. This could not at all times be the case — a really slender interval may at all times be proper. Nevertheless, in apply this trade-off may be very a lot current in all of the fashions I’ve skilled.
Discover the periodic bulges within the chart (round March 10 or April 7, for example). Since we’re producing a 7-day forecast, the bulges signify the elevated uncertainty as we transfer away from the final commentary that the mannequin noticed. In different phrases, a forecast for the following day will probably be much less unsure than a forecast for the day after subsequent, and so forth.
The 128-token mannequin yields very comparable outcomes to the XGBoost forecaster, which had an space 348.28 and a protection of 84.67%. Based mostly on these outcomes, we will say that, with no coaching, Lag-Llama’s efficiency is slightly stable and as much as par with an optimised conventional forecaster.
Lag-Llama’s Github repo comes with a “greatest practices” part with ideas to make use of and fine-tune the mannequin. The authors particularly suggest tuning the context size and the educational charge. We’re going to discover a few of the instructed values for these hyperparameters. The code snippet beneath, which I’ve taken and modified from the authors’ fine-tuning tutorial notebook, exhibits how we will conduct a small grid search:
Within the code above, we loop over context lengths of 32, 64, and 128 tokens, in addition to studying charges of 0.001, 0.001, and 0.005. Inside the loop, we additionally calculate some take a look at metrics: Protection[0.8], Protection[0.9] and Imply Absolute Error of (MAE) Protection. Protection[0.x] measures what number of predictions fall inside their prediction interval. As an illustration, a great mannequin ought to have a Protection[0.8] of round 80%. MAE Protection, however, measures the deviation of the particular protection chances from the nominal protection ranges. Subsequently, a great mannequin in our case ought to be one with a small MAE and coverages of round 80% and 90%, respectively.
One of many fundamental variations with respect to the unique fine-tuning code from the authors is line 46. In that line, the unique code doesn’t embrace a validation set. In my expertise, not together with it meant that each one fashions that I skilled ended up overfitting the coaching knowledge. Alternatively, with a validation set most fashions have been optimised in Epoch 0 and didn’t enhance the validation loss thereafter. With extra knowledge, we may even see much less excessive outcomes.
As soon as skilled, many of the fashions within the loop yield a MAE of 0.5 and coverages of 1 on the take a look at set. Which means the fashions have very broad prediction intervals, however the prediction isn’t very exact. The mannequin that strikes a greater stability is mannequin 6 (counting from 0 to eight within the loop), with the next hyperparameters and metrics:
{'context_length': 128,
'lr': 0.001,
'Protection[0.8]': 0.7142857142857143,
'Protection[0.9]': 0.8571428571428571,
'MAE_Coverage': 0.36666666666666664}
Since that is essentially the most promising mannequin, we’re going to run it by the exams that we’ve with the opposite forecasters.
The chart beneath exhibits the predictions from the fine-tuned mannequin.
One thing that catches the attention in a short time is that prediction intervals are considerably smaller than these from the zero-shot model. The truth is, the interval space is 188.69. With these prediction intervals, the mannequin reaches a protection of 56.67% over the 7-day recursive forecast. Keep in mind that our greatest zero-shot predictions, with a 128-token context, had an space of 399.25, reaching a protection of 84.67%. This implies a 55% discount within the interval space, with solely a 33% lower in protection. Nevertheless, the fine-tuned mannequin is simply too removed from the 80% protection that we’re aiming for, whereas the zero-shot mannequin with 128 tokens wasn’t.
Relating to level forecasting, the MAE of the mannequin is 0.77, which isn’t an enchancment over the zero-shot forecasts and worse than the XGBoost forecaster.
Total, the fine-tuned mannequin leaves doesn’t go away us a great image: it doesn’t do higher than a zero-shot higher at both level of probabilistic forecasting. The authors do counsel that the mannequin can enhance if fine-tuned with extra knowledge, so it might be that our coaching set was not massive sufficient.
To recap, let’s ask once more the query that we set out at first of this weblog: Is Lag-Llama higher at forecasting than XGBoost? For our dataset, the quick reply isn’t any, they’re comparable. The lengthy reply is extra difficult, although. Zero-shot forecasts with a 128-token context size have been on the similar degree as XGBoost when it comes to probabilistic forecasting. Fantastic-tuning Lag-Llama additional decreased the prediction space, making the mannequin’s right forecasts extra exact, albeit at a considerable value when it comes to probabilistc protection. This raises the query of the place the mannequin may get with extra coaching knowledge. However extra knowledge we didn’t have, so we will’t say that Lag-Llama beat XGBoost.
These outcomes inevitably open a broader debate: since one isn’t higher than the opposite when it comes to efficiency, which one ought to we use? On this case, we’d want to think about different variables akin to ease of use, deployment and upkeep and inference prices. Whereas I haven’t formally examined the 2 choices in any of these points, I believe the XGBoost would come out higher. Much less data- and resource-hungry, fairly sturdy to overfitting and time-tested are hard-to-beat traits, and XGBoost has all of them.
However don’t consider me! The code that I used is publicly accessible on this Github repo, so go take a look and run it your self.
