Benchmarking Lag-Llama in opposition to XGBoost
On Hugging Face, there are 20 fashions tagged “time sequence” on the time of writing. Whereas actually not loads (the “text-generation-inference” tag yields 125,950 outcomes), time sequence forecasting with basis fashions is an attention-grabbing sufficient area of interest for large 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 vital fashionable 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 quite a lot of datasets throughout completely different domains. As soon as fine-tuned for particular duties, additionally they declare it to be the perfect general-purpose mannequin of its type. Massive phrases!
On this weblog, I showcase my expertise fine-tuning Lag-Llama, and check its capabilities in opposition to a extra classical machine studying method. Specifically, I benchmark it in opposition to an XGBoost mannequin designed to deal with univariate time sequence knowledge. Gradient boosting algorithms resembling XGBoost are extensively thought of the epitome of “classical” machine studying (versus deep-learning), and have been proven to carry out extraordinarily effectively with tabular knowledge [2]. Due to this fact, 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 can’t go into the main points of the mannequin structure, however the paper is value a learn, as is that this good walk-through by Marco Peixeiro.
The information that I exploit for this train is a 4-year-long sequence of hourly wave heights off the coast of Ribadesella, a city within the Spanish area of Asturias. The sequence is obtainable on the Spanish ports authority knowledge 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 combination the sequence to a day by day degree, taking the max over the 24 observations in every day. The reason being that the ideas that we undergo on this submit are higher illustrated from a barely much less granular standpoint. In any other case, the outcomes develop into very risky in a short time. Due to this fact, our goal variable is the utmost top of the waves recorded in a day, measured in meters.
There are a number of explanation why I selected this sequence: 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 this sort of 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 feel 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 apart from imputing a number of lacking values. The plot under reveals what it seems like after we cut up it into practice, validation and check 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 pocket book.
We’re going to benchmark Lag-Llama in opposition to XGBoost on two univariate forecasting duties: level forecasting and probabilistic forecasting. The 2 duties complement one another: level forecasting provides us a particular, single-number prediction, whereas probabilistic forecasting provides us a confidence area round it. One might say that Lag-Llama was solely skilled for the latter, so we must always concentrate on that one. Whereas that’s true, I imagine that people discover it simpler to know a single quantity than a confidence interval, so I feel the purpose forecast remains to be helpful, even when only for illustrative functions.
There are numerous components that we have to take into account when producing a forecast. A number of the most necessary embody the forecast horizon, the final statement(s) that we feed the mannequin, or how typically we replace the mannequin (if in any respect). Completely different combos of things yield their very own forms 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 dimension of seven days. Because of this 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, nevertheless, shouldn’t be retrained as new knowledge is obtainable. By way of our dataset, which means that 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 Imply Absolute Error (MAE) as efficiency metric. Within the case of probabilistic forecasting, we are going to goal for empirical protection or protection chance of 80%.
The scene is about. Let’s get our arms soiled with the experiments!
Whereas initially not designed for time sequence forecasting, gradient boosting algorithms normally, and XGBoost particularly, could be nice predictors. We simply must feed the algorithm the info in the suitable format. As an illustration, if we need to use three lags of our goal sequence, we will merely create three columns (say, in a pandas dataframe) with the lagged values and voilà! An XGBoost forecaster. Nonetheless, this course of can rapidly develop into onerous, particularly if we intend to make use of many lags. Fortunately for us, the library Skforecast [4] can do that. In actual fact, Skforecast is the one-stop store for creating and testing all kinds of forecasters. I actually can’t advocate it sufficient!
Making a forecaster with Skforecast is fairly easy. We simply must 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 must seek for the perfect variety of lags to incorporate in our mannequin. To try this, Skforecast supplies 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 effectively the XGBoost forecaster does at predicting the following 7 days of most wave heights. The chart under plots the predictions in opposition to the precise values of our check set. We will see that the prediction tends to comply with the overall pattern of the particular knowledge, however it’s removed from good.
To create the predictions depicted above, we have now used Skforecast’s backtesting_forecaster operate, which permits us to judge the mannequin on a check 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. Because of this, 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. Due to this fact, 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 statement as a dummy greatest guess for our forecast, we’d get a MAE of 0.84 (see level 1 of this pocket book). All issues thought of, evidently, thus far, our mannequin is healthier than a easy logical rule, which is a aid!
Probabilistic forecasting
Skforecast permits us to calculate distribution intervals the place the longer term final result is more likely to fall. The library supplies 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 possibly can see extra ends in half 3 of this pocket book.
The concept 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 quite a few occasions, we will assemble an equal variety of various 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 potential forecasts. On this universe, we’d anticipate to see at the least a proportion of the particular values of the forecasted sequence. In our case, we are going to goal 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 check set. The second and third steps are illustrated within the snippet under (the primary one corresponds to the code snippet within the earlier part). Traces 14-17 are the parameters that govern our bootstrap calculation.
The ensuing prediction intervals are depicted within the chart under.
An 84.67% of values within the check set fall inside our prediction intervals, which is simply above our goal of 80%. Whereas this isn’t unhealthy, it might additionally imply that we’re overshooting and our intervals are too large. Consider it this manner: if we stated that tomorrow’s waves could be between 0 and infinity meters excessive, we’d all the time be proper, however the forecast could be ineffective! To get a thought of how large 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 may 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 pocket book 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 contemplate within the forecast. We simply must name the get_llama_predictions operate under:
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 accustomed to 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 under, 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 some extent forecast. If it’s good to study extra in regards to 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 could require differen context lenghts. In our case, we attempt context lenghts of 32, 64 and 128 tokens (lags). The chart under reveals the outcomes of the 64-token mannequin.
Level forecasting
As we stated above, Lag-Llama shouldn’t be 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 will be topic to outliers within the interval. In any case, for our explicit dataset, each choices yield related 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 effectively. These are all greater than the XGBoost forecaster’s, which went right down to 0.64. In actual fact, they’re very near the baseline, dummy mannequin that used the earlier week’s worth as right this moment’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 searching for. 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 attention-grabbing pattern right here: extra protection implies a bigger interval space. This could not all the time be the case — a really slim interval might all the time be proper. Nonetheless, in apply this trade-off could 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, as an example). Since we’re producing a 7-day forecast, the bulges signify the elevated uncertainty as we transfer away from the final statement that the mannequin noticed. In different phrases, a forecast for the following day can be much less unsure than a forecast for the day after subsequent, and so forth.
The 128-token mannequin yields very related outcomes to the XGBoost forecaster, which had an space 348.28 and a protection of 84.67%. Primarily based on these outcomes, we will say that, with no coaching, Lag-Llama’s efficiency is reasonably strong 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 advocate tuning the context size and the training price. We’re going to discover a number of the instructed values for these hyperparameters. The code snippet under, which I’ve taken and modified from the authors’ fine-tuning tutorial pocket book, reveals 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 check 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 very good mannequin ought to have a Protection[0.8] of round 80%. MAE Protection, alternatively, measures the deviation of the particular protection possibilities from the nominal protection ranges. Due to this fact, a very good mannequin in our case must be one with a small MAE and coverages of round 80% and 90%, respectively.
One of many predominant variations with respect to the unique fine-tuning code from the authors is line 46. In that line, the unique code doesn’t embody a validation set. In my expertise, not together with it meant that every 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, a lot of the fashions within the loop yield a MAE of 0.5 and coverages of 1 on the check set. Because of this the fashions have very broad prediction intervals, however the prediction shouldn’t be very exact. The mannequin that strikes a greater steadiness 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 means of the assessments that we have now with the opposite forecasters.
The chart under reveals 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. In actual fact, the interval space is 188.69. With these prediction intervals, the mannequin reaches a protection of 56.67% over the 7-day recursive forecast. Do not forget 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. Nonetheless, 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.
In the case of 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 depart us a very good 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 giant sufficient.
To recap, let’s ask once more the query that we set out initially of this weblog: Is Lag-Llama higher at forecasting than XGBoost? For our dataset, the brief reply isn’t any, they’re related. The lengthy reply is extra sophisticated, 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 diminished the prediction space, making the mannequin’s appropriate forecasts extra exact, albeit at a considerable price when it comes to probabilistc protection. This raises the query of the place the mannequin might 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 shouldn’t be higher than the opposite when it comes to efficiency, which one ought to we use? On this case, we’d want to contemplate different variables resembling ease of use, deployment and upkeep and inference prices. Whereas I haven’t formally examined the 2 choices in any of these features, I think the XGBoost would come out higher. Much less data- and resource-hungry, fairly strong to overfitting and time-tested are hard-to-beat traits, and XGBoost has all of them.
However don’t imagine me! The code that I used is publicly out there on this Github repo, so go take a look and run it your self.