# Gradient-Boosted Trees for Swing-Trade Signals

## Introduction

The application of machine learning to financial time series requires navigating environments characterized by exceptionally low signal-to-noise ratios, non-stationary market regimes, and severe temporal dependencies. Within the quantitative finance landscape, swing trading occupies a unique temporal horizon. Unlike high-frequency trading (HFT), which operates on the microsecond level and is heavily dependent on order book microstructure, and unlike long-term macroeconomic investing, which relies on multi-quarter fundamental shifts, swing trading typically involves holding positions for several days to several weeks. At this intermediate horizon, asset price action is driven by a complex, noisy interplay of momentum, mean reversion, fundamental data releases, and broader market sentiment shifts.

To model these non-linear relationships and extract actionable alpha, quantitative researchers frequently turn to Gradient-Boosted Decision Trees (GBDTs). Over the past decade, algorithmic frameworks such as XGBoost, LightGBM, and CatBoost have become foundational tools in tabular financial modeling, frequently outperforming more complex deep learning architectures. However, the successful deployment of these models relies less on algorithmic selection and more on rigorous data pipeline engineering. The primary challenges in building profitable, generalized swing-trade models are not architectural, but rather methodological. Specifically, practitioners must prevent target leakage from overlapping return horizons, mitigate lookahead bias in cross-validation procedures, and accurately model execution frictions such as slippage and market impact. 

Developing a robust swing-trade signal requires an exhaustive understanding of what works—specifically the inductive biases that make tree-based models highly suited for tabular financial data—and what overfits, which almost universally stems from flawed validation frameworks and a failure to account for market microstructure.

## Efficacy on Financial Tabular Data

Despite the rapid and highly publicized advancements in deep learning, particularly the success of transformer architectures in natural language processing and convolutional networks in computer vision, neural networks (NNs) have historically struggled to consistently outperform GBDTs on structured tabular data. Financial datasets—comprising technical indicators, macroeconomic variables, fundamental ratios, and alternative data—are fundamentally tabular in nature. 

Recent empirical investigations highlight the persistent superiority of tree-based models in this domain. Extensive research by Grinsztajn et al. (NeurIPS 2022) established that tree-based models remain the state-of-the-art for medium-sized tabular datasets, which typically consist of around ten thousand to a few hundred thousand samples [cite: 1, 2, 3, 4]. This performance gap is driven by the differing inductive biases between decision trees and neural networks, which fundamentally alter how each architecture processes noise, feature orientation, and target distribution.

### Robustness to Uninformative Features

Financial datasets are notoriously noisy, frequently containing large numbers of uninformative features. A typical quantitative model might ingest hundreds of technical indicators, such as moving average convergence divergence (MACD), relative strength indices (RSI), stochastic oscillators, and various rolling volatility metrics. At any given moment, the vast majority of these features possess zero predictive power regarding the future price of an asset. 

Gradient-Boosted Decision Trees are inherently robust to uninformative features. Tree-splitting algorithms, which rely on metrics like information gain or Gini impurity, naturally ignore features that do not partition the target variable effectively [cite: 1, 3, 5]. The algorithm evaluates each feature independently and will simply not select an uninformative feature as a split node. In contrast, standard neural networks, such as multi-layer perceptrons (MLPs) and ResNets, process all inputs simultaneously through dense, fully connected layers. Every feature, regardless of its predictive utility, is assigned a weight. Research indicates that an increasing ratio of uninformative features to dataset size actively degrades the performance of neural networks, as the model struggles to zero out the weights of noisy inputs without extensive regularization [cite: 1, 6].

[image delta #1, 0 bytes]

 



### Modeling Irregular Target Functions

Another critical advantage of GBDTs is their ability to model irregular target functions. In financial markets, target variables (such as subsequent multi-day returns) rarely follow a smooth, continuous function relative to the inputs. Minor deviations in input features can correspond to major, discontinuous changes in target values [cite: 7, 8]. For example, an asset's price hovering one cent above a critical technical support level might result in a mild positive return, while a drop of two cents—breaking that support level—might trigger algorithmic stop-losses, resulting in a severe negative return. 

Neural networks, which rely on continuous, differentiable optimization and backpropagation, struggle to fit these high-frequency, discontinuous components [cite: 7, 9]. The smooth activation functions of neural networks tend to interpolate between data points, which blurs the sharp regime changes characteristic of financial data. Decision trees, however, partition data along orthogonal, axis-aligned boundaries. This allows them to easily create step functions that perfectly map to irregular, discontinuous market phenomena [cite: 2, 8].

### Feature Orientation and Distribution Skew

Tabular financial data is generally non-rotationally invariant. In computer vision, an image of a cat remains a cat even if the image is slightly rotated or translated; neural networks are specifically designed to capture these invariances. In tabular data, however, linear combinations of features often destroy the underlying semantic information [cite: 2, 8]. Combining a relative strength index (RSI) with a price-to-earnings (P/E) ratio does not produce a mathematically meaningful new feature. Neural networks are rotationally invariant, which hinders their ability to isolate individual feature thresholds. Decision trees preserve the orientation of the data by splitting on individual features sequentially, which is significantly more effective for financial modeling [cite: 2, 3].

Subsequent research by McElfresh et al. (NeurIPS 2023/2024) expanded upon these findings, conducting the largest tabular data analysis to date across 176 datasets [cite: 1, 9, 10]. While the authors noted that the broader "NN vs. GBDT" debate is sometimes overemphasized, they confirmed that GBDTs consistently outperform neural networks when handling highly skewed and heavy-tailed feature distributions [cite: 1, 10]. Asset returns are notoriously heavy-tailed, frequently exhibiting kurtosis far beyond a normal distribution. GBDTs, which split data based on rank ordering rather than continuous distance metrics, are inherently robust to the extreme outliers that frequently cause catastrophic gradient updates in neural networks.

## Architectural Comparisons of Boosting Frameworks

While Gradient-Boosted Decision Trees share a common ensemble methodology—building sequential weak learners where each subsequent tree attempts to correct the residual errors of the prior trees—the specific implementations diverge significantly. XGBoost, LightGBM, and CatBoost each possess unique tree-building strategies, methods for handling categorical variables, and regularization techniques. Understanding these architectural trade-offs is essential when designing swing-trade models, as latency requirements, data scale, and the risk of overfitting vary drastically depending on the chosen framework.

### XGBoost (eXtreme Gradient Boosting)

Introduced by Tianqi Chen in 2014, XGBoost was the original trailblazer that popularized the gradient boosting framework in competitive data science and quantitative finance [cite: 11, 12]. XGBoost grows decision trees using a level-wise, or depth-wise, strategy. During the training process, the algorithm evaluates all possible splits across all nodes at the current level of the tree before moving to the next, deeper level [cite: 12, 13]. 

This level-wise growth ensures that the resulting trees are highly balanced. In the context of noisy financial data, this balanced growth acts as an implicit form of structural regularization, restricting the algorithm from building deep, highly specific branches that fit exclusively to localized noise. Furthermore, XGBoost natively incorporates L1 (Lasso) and L2 (Ridge) regularization penalties directly into its objective function, explicitly penalizing model complexity and leaf weights [cite: 2, 13, 14]. For swing trading, where the goal is to capture broad, generalizable trends rather than idiosyncratic historical quirks, this strict regularization is highly beneficial. However, the level-wise evaluation of all possible splits is computationally intensive, making XGBoost slower to train on massive, multi-gigabyte tick datasets compared to its modern successors [cite: 12, 15].

### LightGBM (Light Gradient Boosting Machine)

Developed by Microsoft in 2017, LightGBM was engineered specifically to address the computational bottlenecks of XGBoost when scaling to massive datasets [cite: 11, 12]. LightGBM abandons the level-wise approach in favor of a leaf-wise (best-first) tree growth strategy. Instead of evaluating all nodes at a given depth, LightGBM identifies the single leaf node across the entire tree that will yield the maximum reduction in the loss function, and splits it [cite: 12, 13].

This approach allows the model to converge on complex patterns much faster. To handle datasets with millions of rows, LightGBM employs two major innovations: Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB). GOSS accelerates training by retaining all instances with large gradients (which represent under-trained, high-error samples) while randomly down-sampling instances with small gradients [cite: 12, 13]. EFB further optimizes performance by identifying mutually exclusive features—such as heavily zero-padded sparse matrices—and bundling them into a single feature to reduce dimensionality [cite: 12, 13]. 

While LightGBM is exceptionally fast and highly scalable, its leaf-wise growth poses a severe risk of overfitting in financial time series. Because the algorithm relentlessly pursues the maximum reduction in loss, it can quickly generate deep, asymmetrical trees that memorize historical market noise [cite: 12, 13]. When deploying LightGBM for swing-trade signals, quantitative researchers must enforce strict hyperparameter constraints—such as heavily restricting `max_depth` and increasing `min_data_in_leaf`—to prevent the model from capturing spurious correlations.

### CatBoost (Categorical Boosting)

Developed by Yandex in 2017, CatBoost takes a radically different structural approach. CatBoost employs symmetric, or oblivious, decision trees. In a symmetric tree, the exact same splitting criterion (e.g., "RSI > 70") is applied across all nodes at a given depth level [cite: 12, 16, 17]. This rigid structure ensures perfectly balanced trees, which greatly reduces the model's propensity to overfit. Furthermore, oblivious trees allow for exceptionally fast inference times during live trading, as the tree structure can be evaluated as a simple, flat array rather than a complex branching logic [cite: 12, 15, 18].

CatBoost's namesake feature is its native, sophisticated handling of categorical variables. Traditional GBDTs require categorical data (such as sector classifications, regime states, or macroeconomic sentiment labels) to be pre-processed using one-hot encoding or integer mapping. CatBoost utilizes a proprietary method called target-based statistics, which calculates the expected target value for each category dynamically during training, effectively transforming categorical inputs into continuous, highly predictive numeric features without exploding the dimensionality of the dataset [cite: 13, 14, 19].

Crucially for financial applications, CatBoost implements a technique known as "ordered boosting." Traditional gradient boosting calculates the residual error for a data point using a model that was trained on that exact same data point, leading to biased gradient estimates and target leakage [cite: 13, 16, 17]. Ordered boosting mitigates this by maintaining a randomized permutation of the training dataset. For each data point, it computes the residual error using only a model trained on the preceding data points in the permutation [cite: 13, 16, 17]. This algorithmic defense against target leakage makes CatBoost highly resilient when modeling the inherently autocorrelated nature of financial asset returns.

### Summary of Architectural Trade-Offs

The choice of boosting framework ultimately depends on the specific constraints of the trading strategy, the scale of the data, and the risk tolerance for overfitting. 

| Feature / Metric | XGBoost | LightGBM | CatBoost |
| :--- | :--- | :--- | :--- |
| **Tree Growth Strategy** | Level-wise (Depth-first) | Leaf-wise (Best-first) | Symmetric / Oblivious (Level-wise) |
| **Overfitting Risk** | Moderate (Controlled via explicit L1/L2 penalties) | High (Prone to deep, asymmetric branches) | Low (Symmetric trees and ordered boosting) |
| **Training Speed** | Moderate | Very Fast (Leverages GOSS & EFB) | Slower on CPU (Highly optimized for GPU) |
| **Inference Speed** | Moderate | Fast | Very Fast (Due to flat oblivious structure) |
| **Categorical Handling** | Requires manual encoding (e.g., One-Hot) | Requires integer mapping | Native handling via Target Statistics |

## Mechanisms of Feature and Target Leakage

In financial machine learning, the failure of an algorithm in live trading is rarely due to a deficiency in the model's underlying mathematics. Instead, the failure is almost universally a result of a compromised data pipeline. Gradient-boosted trees are highly capable optimization engines; if they are presented with an opportunity to "cheat," they will ruthlessly exploit it. 

Data leakage occurs when a model is trained using information that would not have been available at the exact point in time a live trading decision is made [cite: 20, 21]. When leakage is present, the model effectively learns to recognize outcomes that have already occurred rather than predicting future risk [cite: 21]. This results in models that display spectacular, near-perfect backtest metrics (such as inflated Sharpe ratios and low Root Mean Square Errors) but fail catastrophically the moment they are deployed in production [cite: 21, 22, 23].

### Direct and Implicit Feature Leakage

Direct feature leakage occurs when a proxy for the target variable is inadvertently included in the predictive feature set [cite: 24, 25]. In quantitative finance, the most common source of direct leakage is the use of revised macroeconomic data. Government agencies frequently release macroeconomic indicators (such as GDP growth, inflation figures, or unemployment rates) and subsequently revise those numbers weeks or months later. If a researcher trains a model using the finalized, revised historical data, they are injecting future information into the past. At the time a historical trade would have been placed, only the initial, unrevised "point-in-time" data was known [cite: 20, 24]. A model trained on revised data will appear highly prescient in backtesting, but will collapse when fed real-time, unrevised data feeds in production.

Implicit, or combinational, leakage is significantly more insidious. In these cases, no single feature contains explicit future information, but when multiple features are combined by the non-linear interactions of a decision tree, they indirectly reveal the outcome [cite: 21]. 

Another frequent pipeline error that compromises otherwise robust models is premature featurization [cite: 25]. In an effort to clean the dataset, a data scientist might apply normalization techniques (such as Z-score standardization, min-max scaling, or Principal Component Analysis) across the entire historical dataset simultaneously. By calculating the global mean and standard deviation across the entire time series, the statistical properties of the future data are leaked into the past. When the data is subsequently split into training and testing sets, the training set has already been normalized using parameters influenced by the test set [cite: 22, 24]. To prevent this, all scaling, imputation, and feature engineering transformations must be strictly fit exclusively on the dynamic training window and then subsequently applied to the validation window [cite: 20, 22, 23].

### Target Overlap in Multi-Day Horizons

The most pervasive form of leakage in swing trading models stems from target formulation. Unlike high-frequency algorithms that predict single-period (e.g., one-second or one-minute) returns, swing-trade models attempt to forecast multi-period events, such as a 5-day, 10-day, or 21-day forward return horizon [cite: 20, 26].

When calculating a multi-day rolling return, the target variables of chronologically adjacent observations overlap heavily [cite: 20, 26]. Consider a model designed to predict a 5-day forward return. The target calculated on Monday represents the price movement from Monday to Friday. The target calculated on Tuesday represents the price movement from Tuesday to the following Monday. These two target variables share four overlapping days of exact market trajectory. 

If Monday's observation is randomly assigned to the training set and Tuesday's observation is assigned to the testing set, the model will learn the specific, idiosyncratic market movements of the testing period during its training phase [cite: 20, 26]. Because financial returns are highly serially correlated, the model effectively absorbs the future test set data. This phenomenon, known as overlapping target leakage, leads to massive overestimations of out-of-sample performance if the cross-validation methodology is not explicitly designed to handle it.

## Time-Series Validation Frameworks

The integrity of a swing-trading model rests entirely upon the validation framework used to assess its performance. Because financial time series data violates the independent and identically distributed (IID) assumption required by traditional statistical testing, standard machine learning validation techniques fail completely, producing overly optimistic and ultimately dangerous models.

### The Failure of Standard K-Fold and Walk-Forward Validation

Standard K-Fold cross-validation, the default methodology in generic machine learning, randomly partitions data into *k* folds without any regard for temporal ordering. By shuffling the dataset, K-Fold allows future information to heavily contaminate the training set [cite: 27, 28, 29]. A model evaluating the year 2015 might have been trained on data from 2018; it has essentially seen the future. Consequently, standard K-Fold CV is categorically invalid for financial time series analysis and guarantees severe overfitting [cite: 20, 28].

To preserve chronological integrity, quantitative researchers historically adopted Walk-Forward (also known as expanding window or rolling window) validation. In a walk-forward setup, the model is trained on a defined historical period and evaluated on the immediate, subsequent chronological period. The window then rolls forward in time, repeating the process [cite: 30, 31, 32]. This approach strictly prevents lookahead bias, ensuring the model never sees future data during training.

However, walk-forward validation suffers from a critical flaw: it evaluates the model on only a single historical path [cite: 26, 28, 29, 33]. Financial markets are highly non-stationary; the economic regime of the 2008 financial crisis differs wildly from the quantitative easing regime of the 2010s. By evaluating performance along a single chronological trajectory, the resulting metrics (such as the Sharpe ratio) exhibit extremely high variance [cite: 29, 34]. A strategy might show excellent theoretical returns simply because its parameters happened to align perfectly with the specific sequence of historical events in that single path [cite: 28, 29]. This leads researchers to optimize hyperparameters that fit the noise of the specific historical sequence, failing to build a robust model capable of handling unseen future regimes.

### Combinatorial Purged Cross-Validation (CPCV)

To resolve the path-dependency of walk-forward testing while strictly enforcing the chronological boundaries necessary to prevent target leakage, Marcos López de Prado introduced Combinatorial Purged Cross-Validation (CPCV) [cite: 29, 30]. CPCV elevates cross-validation into a comprehensive framework capable of generating rigorous statistical inferences for financial datasets.

The CPCV methodology begins by partitioning the time series into $N$ sequential, non-overlapping chronological groups. The algorithm then selects all possible combinations of $k$ groups to serve as the out-of-sample test sets, with the remaining $N - k$ groups forming the training set for each split [cite: 29, 30]. By combinatorially selecting different periods of history to act as the test set, CPCV generates a multitude of distinct out-of-sample backtesting paths. 

The primary advantage of CPCV is that it does not output a single performance score. Instead, it generates an empirical distribution of out-of-sample outcomes [cite: 28, 29, 35]. Rather than seeing a single Sharpe ratio of 1.84, a researcher analyzing the CPCV output might observe a Sharpe distribution ranging from 0.6 to 1.9, with a median of 1.1 [cite: 35]. This distribution reveals the true stability of the strategy across a diverse set of plausible historical paths [cite: 28, 29]. A tight distribution with low variance indicates a genuine edge robust to regime shifts, whereas a wide distribution suggests the model is highly brittle and merely fitting to noise [cite: 28, 35].

### The Mechanics of Purging and Embargoing

Because CPCV constructs test sets from combinations of chronological groups, a test set will frequently be preceded by, or followed by, a training set. To ensure that the overlapping multi-day targets discussed previously do not cause leakage, CPCV integrates two mandatory, strictly enforced boundary mechanisms: Purging and Embargoing [cite: 29, 30, 33].

1. **Purging:** Whenever a training set chronologically precedes a testing set, the algorithm actively scans the boundary between the two. Any training observation whose target calculation window (e.g., its 5-day forward return horizon) overlaps with the beginning of the testing set is permanently deleted from the training data [cite: 20, 29, 30, 33]. This strict temporal separation ensures that the outcome of the final training sample is completely resolved before the first testing sample begins, mathematically severing the link of target leakage [cite: 20, 29].
2. **Embargoing:** Financial markets exhibit persistent autoregressive effects and feature memory; a volatility spike on Tuesday is highly predictive of volatility on Wednesday. When a testing set precedes a training set, target overlap is not an issue (as the test target resolves before the training target begins). However, due to serial correlation, the market conditions immediately following a test set are highly contaminated by the test set itself [cite: 20, 30, 33]. To prevent the model from learning these immediately adjacent correlations, an "embargo" protocol actively discards a specified buffer duration of training data immediately following the conclusion of the testing set [cite: 20, 26, 29, 30].

[image delta #2, 0 bytes]





| Validation Methodology | Preserves Temporal Order | Handles Overlapping Labels | Backtest Output | Risk of Overfitting |
| :--- | :--- | :--- | :--- | :--- |
| **Standard K-Fold** | No | No (Massive Leakage) | Single Aggregated Metric | Extreme (Invalid for time series) |
| **Walk-Forward (Rolling)** | Yes | Requires manual shifting | Single Historical Path | High (Severe Path Dependency) |
| **Purged K-Fold** | Yes | Yes (Via Purging/Embargo) | Single Historical Path | Moderate |
| **Combinatorial Purged CV** | Yes | Yes (Via Purging/Embargo) | Distribution of Paths | Low (Reveals Variance and Stability) |

## Execution Frictions and Market Microstructure

Assuming a quantitative researcher has successfully engineered an XGBoost or LightGBM model using point-in-time features and rigorously validated it through Combinatorial Purged Cross-Validation, they may still find that the model fails to generate real-world profits. A persistent gap between an algorithm's theoretical out-of-sample performance and its live trading returns is almost entirely driven by the failure to accurately model execution frictions and market microstructure.

Algorithms optimized solely for abstract predictive accuracy—such as minimizing Root Mean Square Error (RMSE) or maximizing Area Under the Curve (AUC)—will frequently generate high-frequency, low-conviction signals. If a swing-trade model operates with excessive turnover, attempting to capture minor basis-point anomalies, transaction costs will rapidly erode the entire statistical edge [cite: 36, 37]. 

### The Components of Trading Costs

Transaction costs in swing trading consist of two primary components: explicit costs and implicit costs [cite: 37, 38]. Explicit costs are deterministic and relatively transparent. They include brokerage commissions, exchange clearing fees, and regional regulatory taxes (such as Stamp Duty in the UK or the Securities Transaction Tax in certain emerging markets) [cite: 36, 38]. Because quantitative strategies execute high volumes, explicit per-trade costs are generally negotiated down to fractional levels, but they remain a constant, unavoidable drag on equity over large sample sizes [cite: 39].

Implicit costs, however, are dynamic, difficult to quantify, and scale non-linearly with market volatility, asset liquidity, and the size of the order. Implicit costs frequently outweigh explicit fees and consist of three main factors:
1. **Bid-Ask Spread:** This is the foundational cost of crossing the market to execute an immediate trade. At a minimum, a market order is immediately out of the money by half the amount of the quoted spread [cite: 37, 40]. In periods of high volatility, spreads widen considerably, punishing high-turnover models.
2. **Slippage:** Slippage represents the price difference between the theoretical time a trading signal is generated by the algorithm and the time the order is successfully matched and filled at the exchange [cite: 36, 38, 39, 40]. Delays in the signal-to-execution pipeline, thin liquidity, or rapid price movement during fundamental news events lead to severe execution slippage [cite: 36, 38, 40].
3. **Market Impact:** This represents the degree to which a trader's own order moves the market price unfavorably [cite: 37, 38, 39]. Large block orders deplete the available liquidity resting in the order book, pushing the average fill price further away from the theoretical entry point [cite: 37, 40]. 

### Transaction Cost Modeling in Validation

For multi-day or multi-week swing strategies, a single tick of slippage is less catastrophic than it is for intra-day high-frequency trading [cite: 36]. The broader price movement of a multi-day trend usually absorbs minor entry inefficiencies. However, cumulative slippage over hundreds of trades over several months will significantly depress the strategy's risk-adjusted returns.

To accurately evaluate a GBDT swing-trade signal, researchers cannot rely on "flat" transaction cost models. A flat model applies a fixed fee per trade, ignoring the reality of the order book. Flat models fail to capture the dynamics of slippage or market impact and are categorically invalid for assessing quantitative viability [cite: 38, 39]. 

Instead, robust execution modeling requires implementing piecewise linear or quadratic transaction cost functions within the backtesting engine. In a quadratic model, the cost of execution scales exponentially with the order size relative to the asset's average daily volume [cite: 38, 39]. If quadratic execution costs are not penalized directly within the model's objective function, or applied strictly during the backtesting phase, the GBDT algorithm will inadvertently over-allocate capital to illiquid, micro-cap assets. These assets often display high theoretical predictability in a vacuum, but the liquidity is so thin that any attempt to trade the signal at scale will result in market impact costs that immediately destroy the alpha.

## Conclusion

Developing profitable swing-trade signals using Gradient-Boosted Decision Trees requires balancing the algorithm's powerful non-linear capabilities with strict methodological controls against overfitting. The empirical evidence clearly demonstrates that tree-based architectures—such as XGBoost, LightGBM, and CatBoost—are fundamentally superior to standard neural networks when modeling the highly irregular, skewed, and uninformative tabular data typical of financial markets. Their axis-aligned splitting logic naturally filters out noise and prevents the catastrophic degradation that dense networks suffer when presented with uninformative indicators.

However, extracting a genuine edge requires looking beyond the architecture and focusing entirely on the data pipeline. The tendency of models to overfit is rarely a failure of the boosting algorithm itself, but rather a consequence of target leakage, premature featurization, and flawed validation frameworks. Standard cross-validation and walk-forward methods introduce severe lookahead bias and path-dependency, rewarding models that memorize specific historical quirks. Quantitative researchers must adopt rigorous, chronology-respecting frameworks like Combinatorial Purged Cross-Validation, actively applying purging and embargoing protocols to sever the temporal correlations that lead to phantom out-of-sample performance. Ultimately, a machine learning strategy is only as robust as its integration of market realities; swing-trade models must explicitly account for quadratic transaction costs, slippage, and market impact to ensure that theoretical signals successfully translate into realized profitability in live markets.

**Sources:**
1. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF1BU4UYJzdJSBk-gzP2vsnL6DH_Aelv9vjPER97wwOPZs9jBpVNR9o-KAIgzBzxN9iFnCxBVChMaiWwB1y3S0_bL_8gImVshLAckaz4q381V9wDuvz)
2. [mdpi.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGpguqC7O9izAdGJRewmbS2_IKaFnGuTDkcf8l84bZ1FL2mNRmWzWBEPFZMrfqqH1P_pFU3w8dx3vP_op4ncnhXvKuHwff3Q-Fk9ONRATxU-cWr9TWtTNyL5J0=)
3. [neurips.cc](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE0IC0RyC61WRhcsvIwK0CW4IT553ECxaLAUP3KDsh1M9cI3Iqmc6rDngB6iR4FoO4ZKtEeM99qira4V3CHNWQ6xB_C-Nr5sTSYJScN3B6TBifkKh9L4lCU0YlSaMeMfQjG)
4. [scribd.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHdMae-qEukpGacVdB-tZ2Lp-HYSg1yMxUxh5SRehZGC1bQLEZHv2OB58yN7c_bbIhbU7AUGfJOIOAmdeIErWJNgfK4zfVUA6an1bWL4hc9ukZcS8DmuFznj3thm4IaDCjCwaPhr4T6MCrNYNZDrdPE8iD-nCndKBJIAfS0IFO_T9oLcqjzFRHkhjY0L_yW-Ui8DQI_1Z5ZNeS_OW-iKknbWIwHerR6Bo1LZLRW1-FlTqgk1zBGY8EmIx3FAMkDMsfX9vGLCYJd3z_7ytnpkk01A916YlS6XRne3Bzj96Gx73A=)
5. [mlr.press](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFK0_YOtL1ECFvn4kotJNfN_AyXRbHMlgZnWidT9evSGDY9UGsQxhksjmrMiTinMviG2QvUU_HYwnqoOCrufsbzMHwErJH5q5phAO8G5rDI-VyFsHA-Hurmv1Anq7hmVRTLN5KTKj0g7oZfX07DO9Q=)
6. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEoOMXV2ODywK-f2mm8QOGOAVyjMnDMpnHM1GA77UIVybohxBkOdAdooktPF6BNoDMrbIwvf17wkyoiKuQgOf-eMFDThfaSvHEdjATy1YJzEsEWIj16iBZPawqRpvRaHeIDAyT2J_QzhVqaixpd4XR30l9F8wmfc-1fPdkkh1lFiLGkj2EU8p7iuIhdK3IIv7gpui3lRv1rLFN82DpScEONGoMcO6gNSVOljHAdWyv_l98t2bozVCa6)
7. [aaai.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGyrXJRSJdD-xjj2d_2_oVy-9hTCDdxhX8iyOCkDvL6vTYatMXNEhn8opijXV4_ajwJpChKrJDXHj9Od3QsydymCSSvhYZH1oBp8Wh4kRCK-bNvdogzunUu9Iux3S6C4M6ja9Mh9bBK08m8kRIsk-LZW4dst7Q=)
8. [github.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFtUfm6_FNGCK1FokvyzNPSeHewsAqkjBjgHNvPN_G0rsZV4yH0B3ON087b65COl0AwYSrQNnfHOp3mkoErKdkL-rKEx3CuN8VZFiI-K0vE2gT8h8NlpceYdJNe6KurYd50Pl6H50oKapcRzpnHwU27BTx4UE4uxwi1cg==)
9. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFCfG6R_iNTqfot0sMQEA4BLYA2K9VpENMaOjPUZKUuyO96P7y9c9dB9UI4N9g7BT80qWqSeRziXwMSs3wbcIOO3DRqvZ1wRVtZxcDotQXrPLD48aEjBjsE)
10. [huggingface.co](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG779g83RiD0BMAOsCSI7JUgcjlGhQOOPnVzptUUKDSrzh4sJHBpK6gt0TiQTmzVkSRBvsNXp0jhLcSZdis7k2Prh9CMFtHzJ_vN2WWRRYaU4rhqe6nNNYLKjMIYCd9xydl0uo1RIlbsu6ugqO2EVE=)
11. [mdpi.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH-ivV91tHIzNkyXywT_VxVyodwa2E0OwAj6scIJLjiy04e4_6Po3uBAkjfyzlskVP8hgTnhBJMfJVurPc5Vpr0OIj9ASdBCQdFtjFW8ox72axUpSjdmxUSQpK1XwpZ)
12. [apxml.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFVqzxi_j8JeGeXwzP8qtSB13REkJbLkf6YTXYfKhi9XOkbYpHn271LIWB0vbetpLSOhmjY1jTEe9pQHG8RBjiIXlWDxS6pxq_7uxALjWr0faSkgyJPAdhwvCsgrvm4d9-T9-IYCKXTVJYYo0Q=)
13. [createbytes.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQENSHt87gOD277MfkzCb5e3k0MR3xHW9CT0koJOZ_wWW3WbEitHxaeSkVWKpzEM6P0VZW2yTQ6RqR8rMLCCS_mfRgzmBSHY4-mZWbNNlbTu-Oc2NmjqY-kBnvb8QQLg2MgQL_Twi9cIAXYj4kMFTidpUJjqdVbWEtSqLkd4iETTZOc=)
14. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEpECUEFrVpmxGHXoAyr8g8Jx5vWeNhwbtof9ajnRj4ZF1sK-4X077L6ei6Xq-0-qcVAmQjE8RCbtH-giGW4kifKzejHwoZu5N_MhCKU2l--CgQ9hvjvIRWxrJoSr8IOarAhGWSmfRcBXcgyKUr-FOl3n3dZfOt2K0dR2gwTG6jDMDO3Ml6vnE_nIgUG5jdxuezjJ36gTKCYxcVBDD8Nm3D)
15. [joinplank.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH2UlyYqDeFUcU2QBJxwdGdK17stMKYXffDgecooilrHjLqdGe6QKxU1WjMV7l0CEfF327-EDf9QVbrsRyCZBviyfOxNgr6-IHKCCqFNwRT43bfWJ1pT_PpidM7EvZJkZimx1xHXAZAZgAC0dL7yEKlJg==)
16. [mdpi.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHbycMpvFE-Rem05qgRr0Ms6zuISXml2q5wBv78mJQGWOG6zU7KhqxnkOJZieBOgI1l7jd5_HiTLHyzsOEHoNZzSEw_fe_EC-bZnyNnuu68-qt2iB7sZ5w1ZjKa5sc=)
17. [preprints.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGBGPzC_dp0jhBxNdFwbpYxn3WejpPQnyq3cIsMipYWvhrL2NeahvBMaFo-XZ1yemL-XyD1O1u2_k6a5svuLnX4X-C4LANJ0g-qf7VmHdqAwo3pC2Kfw86XH9_1myxY1tfeggwW7A==)
18. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHKLsfnh4S6th7wTsGt2MNt4EM09OAK5URdjDXXpfL61kz9Ici9faffScoda6IVpF_HD9aYgnxKEMfuBbLRJWAllrx5gHXoRfS9T6mz4hReDI3tTD4GZjeRTVSk_802MItrcW94ErAi)
19. [puc-rio.br](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHnt81sf8tyX2gU1loIYxb_QKkJzF4poM-MtAAtDpDoOSvbEo4_PWLW6kc1ELNO91vL0ksb5ujGibKjcClCPndpSqwEKc9CzJSYHl9PBVZXMN9Lxm7CDSbelXNjLV5Vr3wJYpY8CBV4e7k3j-Vw05NEqPRBsHnD9abAt1ojVXQkMrcrwEd2e4_Z7-Z8NHP0AJWWe49Dm1ASaznZXA==)
20. [ijqqr.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFmgg-7k3P7RG_RzxR4s5wkCjmBAAZVvf5ugDtMWbuXK-IJQ72SJvXPGrP_0fEtSZ_e87KoRyBS1aam8hlMA3jWKH5uwplADVcs6UAqEChfEYpSDMyS9TTquSchJux4_zJh5j7Hmo7msXNnDyZomn7kOQ==)
21. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGcABKJCa2HYYwyMQFKrn2-hm50q1ed1NtqTAQ1HVeTeOXzNiI-kZlFnyteeT_2A4t8fyXp-LIyVJpXun-mSwItk8BnMsWPp2K6AnyWj8DsuViWP-zT4eJBmd3V9oIPQ0tlvYXln32CWTPdIGWFImBWvXDbPoteEMXOPCquLo_MKpbfO8hRG_VXnkbZvYcbnm_JQEYQfESplzuftrygnheysYvFkw==)
22. [ibm.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFpAoLHexaOlvdQkMeUmf8EwFsbn_Lt5SXhkjC77wkSmBw7nEY1yp2KwLlLgyKrA10T0ZHOj9PGw8yB61gzX6wkCl7v3OC9QubHPfC-FpvzJjs6jWcWiKT5icP-hojir4nz76C9vfIs-VVFzqRGU45gWvmE)
23. [machinelearningmastery.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG8mYmI0S7BHTVcXuSOS6WMMemt8kQ0gaPj5P81cdQGd3oGcXVQVzgbYK0_bW7qBloLUC6Goi_3iId1wSIZAerUAjAdkwI_J58sN2_9OZuuHBRWn1hdm6foAZZBbvHvhuEJuwjS5UazNclvP9QkOMhGSV_v8O3K)
24. [rayidghani.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHfV2BV6hDkJmOBr_HdJAjq0a-RL1VNUkEQTvNN2vefnzWbBz0zlJCiLrMmE_LgSabLrf5nBOCnrp8G3ldSWb_SbzGnfdKzQyePX2tbocpRM4AQqmVS-t_TfQjxGrhW63pCjXppmGewKUux0WDZJkuk1vuRGwwyqLqxJ57qUT8xBcsnaKToDXDo-ikxQbx7)
25. [wikipedia.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFp1RlWu9QG30i-FGXNtxCA1Ui_tOGJkTqDFnBLTOLNkk1g_qSXm_-G334MVJ31ckAF8ae8G-aIsXN-tHRe3_gW6iin1mm3qwVDkiPRQWBGdl3ysJRqaBx_YNPRkUI18ydQjcZYaPw342wMFXIs)
26. [quantinsti.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFmhgWcAzrfZLsAC13DDGDkqKN7eJW1PG6a5LexLqaeTAcJrgfHeTbl2iG0N_m4SURSIbDtejgtZB17pGCjVCOotGvEKxOSkafv5Z_VTjxLzN--5DB_ocjP_bFAtmOkAaBofRsa9NExCq2K-Egwkr5HKyIcpnQChT-K9fSDWH-BCQ==)
27. [quantbeckman.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFbGKwLdVHi-Cemd5JlDaKuZHzMWtn4l6jzEZOWKDw3L_MgFmeWxbXPYRUQmRexowM7jU6RYxbf9W0yS7SS4FeZBP1i3ewngB4W99WHJM5-8L8M5_Pc0v8-xsiH5ZEQD-gZlqUu_VBRXJ05-x_3WSzKoVHDUtBA29uUIA==)
28. [quantbeckman.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFYXXWX6m29KnUvZHHXnsGD7byaCRM5wiVBQRqRKfKOOWFFyXkNPO_03dH5a_iTlWguIBvIo4HBZZjHtg4SF86oTsLH3IkH0CrRYTeGyDyiOMsttyvof520MJT-GuZg5FE7SgyJUPHx5gX0mUxNsWxQ7uSM9p5d7tE=)
29. [wikipedia.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGvcFfXv08DwjPTU3lToI4iksn1bFkSuggxWUUKJT1T3YXDZw8I9i4ul3mctfvftC7qb2CHLbd0sQyUrIjYhcRMBqlqcqK8se12XkKF28lc45tDkPvmzF7DDOGYmd0Ty-RhZc7XUPp59DZz)
30. [grokipedia.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH4JqPcaExT7zfPbSG6RmzOWVKp1U34rfFkwQ08QHCg3IiKlQW0Ir_G5dGammoVuOLIDcNnldE9kqOmtjRKOYRO5zlHUVDiN-R_mLofoF2X2iMa7eKHnHfO-m_v1QBg-LZzEoRxXORR4Q==)
31. [gitconnected.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF6vlHF-6tPsUvznjIIBp1QGMm7wmDwFH6TIB8vsT7eB5sUH-DpJiTob0NC0r87VnEc6ZGQ554SdgagMttrWfxusgTh5ncV9MAcVcfQBnNWU4bbxjOoQrEU7ksmvGiV33gkcYDmVGCXZZ_CaS23F2E1K7LSKK3lIYDkEDoD0Nso3d9Hw2ZhHx6y68_B6JHgdzmeIj_5Th8wWH9E1c1lCj8HXypGb1s7xEXC3g==)
32. [dndlab.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEwqMNZUzAReaz3-Dp-C_w85bqL8S0QkgIaefcccBtiDkhGiUKNb4XrY9pln-DJZftPP72MxzRTK_q_98tXxvTFJFJFQrbgOhlcWvSkvUlamp9hxPTqO4NQaBNcZWNf4DbRx_zokpYMLMdZc1fgdKTCcBl373s8DpgmnISzGLBWNQpa3_IH9dwTyMPurvFesysOQp8jNYcN9LuC_S_DCFozekUmpg6Rzg==)
33. [github.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGnZ4t30Tex3olqFc9pClagZWpHmtsDvANU1PZdDuF4lmsHZk_TGfjB4FlysTkfspcafiXaK3ZpPiEivtl6TdIzL3JTYtwGw1InU7TizKiyT0ymyw58BEtpMIwfFHjBZ-4lqdW3yJw=)
34. [towardsai.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGSlnLXIpCUmCFoBUukG2WqHfiMj8EpqvhTVBYcuGQAxVqF5a2c4fm-Bbh7olAWKI3SVhKkUya-JQgMTKy38wtSAtV97mFKtylvn-QE2fYXCDsMswLeAHvBOKQ2PhqVX9st5YGgLtAyHCbjSFw79kcpf9QyPBKYWfGOp7zSxnycNcIg_9hSaOjGv71JTwUIc-P_rzQy9bDRI50=)
35. [reddit.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH1eS8-vyLUPPxAHe4Nt__zZreogev3q9tQIM2fdpTTRpeWd4zmnbv0twRBRdNkeQR7eSAYjr8EkQQDFX_DIQaEJ3W-iB93ETPMmVu7WKHxdgLpiypE7hacd34=)
36. [stratzy.in](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGGvYLSjuHYBFNIs9gXSIL5kQVX8--3wBykIBFTxBKBDEbriP2rlRV6PT8HxQqZtA8ocMg1eg4U_q4CoO-VaurrpuKFsF4DseHQN4p9VXSO8sG15fU0KH-16Q==)
37. [grahamcapital.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHAFdrndWLE-11-LwQZevGDa78zLPESvyH3qGyR4N7P4uVNcIH2wrGsxDdXcFfBYfToLVQ3zFtewL42latdxoAti5H12vlnDanU7eczZxYd5BIKP5MFFqReqEzN2tPa1g1y2XyEa6QNCDpc-E6caoJCfWjDn3lDIUrQT3uSJ50Oz4cUCq5PWL8HAzNVlGIIxvnzV9FnkXkQXSRUIWY=)
38. [quantstart.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE2rAfrD7cR-_Mrur9zNn1VJb3wx2vvoEvkUzgqbqh1_3AGIoKbiMLHXU2lj6zChwfkUMXTpFMcxx3sGFHg4Ql01ucCM7segSHj-opHCL7W1PxO_A-hg6PHLkVKy1LpE7huukQ7BFa23qV9rvDosWGz4kfYK3zksTpstK7pMX1dQsnMICP1bEZfqWxlPGiq_4jzJDqsndgmZhxz)
39. [bsic.it](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF3ZIVBVf57AsjZUORPFzLZ3QsZ344f2jQj7pnV9EF4U8uKS5CXDbXRqOBB9MuNpz_W0DpslDXRjLNdbWrC1I8AEeBp2KgXIf33zxuu761n-iMdn6sFn6fJ4kwArMe4aukhCOJTsAWJYwsIznANJxgoqy4PzrxICCIxXjdbzw==)
40. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHhKss8GLiG3M4KvRfzZclU1JKVrpk10s-mGUTlCGsuhRkwFJebnGMGDse6bjHNYxnL3Pf4FdppT2WNZIhE0kcLi9ZgC_1MEJfPp9w8NeoiQtkur2IUVc5LAoSEnqnrSoz2YwNa9N4Ej6SPtWXzyv9m3634gKgke3ZhFy8CL84cuxtNhvw1lhfhCYytn-y9229XZHrzKe0ukNc5G8bI6RyEA94nFTy-q_whQuTH4pgj5ysfZXJW8tcSWA==)