Pith. sign in

REVIEW 4 major objections 6 minor 36 references

City-Level Foreign Direct Investment Prediction with Tabular Learning on Judicial Data

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper contends that 380 judicial-performance indicators computed from twelve million court judgments can predict city-level foreign direct investment with R2 at least 0.92, outperforming ten tabular-learning baselines on both…

desk verdict Useful applied paper with a serious methodological question mark: the judicial-data pipeline is genuinely new and the benchmark is broad, but as written the column encoder leaks test information and the headline 0.92 R2 is not a clean predictive claim. read the letter →

arxiv 2507.05651 v1 pith:NX33GWBO submitted 2025-07-08 cs.AI

Tianxing Wu , Lizhe Cao , Shuang Wang , Jiming Wang , Shutong Zhu , Yerong Wu , Yuqing Feng This is my paper
classification cs.AI
keywords foreigndirectinvestmentpredictionjudicialperformanceindicatorstabularlearningmixtureofexpertsarithmeticattentionadjudicationdocumentscross-citycross-time
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that official economic statistics, which can be manipulated, are not needed to forecast how much foreign direct investment a city will attract. Instead, it builds 380 indicators of local judicial performance from more than twelve million public Chinese court judgments, forms them into a table of city-year rows, and fits a new tabular-learning model to the FDI totals. The central result is that this judicial-data-only route predicts FDI with an R2 of at least 0.92 on both a mixed-year cross-city task and a cross-time forecast, beating ten existing tabular learners. If valid, this gives local governments a transparent, verifiable, and manipulation-resistant input for investment decisions.

What carries the argument

The machinery is a tabular dataset of city-year rows (265+ rows per year, 380 columns) plus TLJD, a model built from it. The model first encodes each indicator value with a row encoder (a separate linear map per indicator) and a column encoder that min-max normalizes a whole column and runs it through two multilayer perceptrons, then fuses the two representations by element-wise product. Transformer layers with arithmetic attention—additive plus multiplicative multi-head attention—let each indicator embedding draw context from the other 379 indicators. Finally a mixture-of-experts head, with one transformer expert per indicator type (procedural justice, distributive justice, judicial efficiency, judicial characteristics) and a gating network, weights the four expert opinions per city to produce the FDI prediction, trained with a regression loss plus an expert-responsibility loss.

What would settle it

Recompute the mixed-year CCP results with the column encoder's min-max scaling and MLP parameters derived only from training rows, holding test rows out of any column statistic, and compare the resulting test R2 to the reported 0.9217. If the score falls toward the w/o-ce ablation's 0.8821 or lower, the reported advantage of the column encoder is an artifact of using test information during encoding.

Watch

Extended reading notes

Core claim

The paper's central claim is that judicial performance, measured by 380 indicators derived from over twelve million publicly available adjudication documents, carries enough information to predict a city's annual FDI, and that a model built for tabular judicial data (TLJD) makes this prediction at an R2 of 0.9217 on the mixed-year cross-city task and 0.9242 on the cross-time task. In both settings TLJD reports the best numbers among ten baselines on every evaluation metric. The underlying assertion is that what courts actually do, not official GDP statistics, steers or signals where foreign investment will go, so a manipulation-resistant input can replace unreliable official economic data.

Load-bearing premise

The method assumes the column encoder, which computes statistics over all rows of a column, is fit without using test-row information; if test rows influence the features of training rows, the reported R2 values are inflated and a single new city-year cannot be scored independently.

Editorial extensions

If this is right

  • Local governments can estimate missing historical FDI figures from court records alone, giving them a check on official numbers that may have been manipulated.
  • The same pipeline produces next-year FDI forecasts before official year-end totals are published, because judicial data is available as soon as judgments are issued.
  • The gating weights identify which aspect of judicial performance matters most in a region—procedural justice in high-FDI cities, distributive justice in lower-FDI cities—offering a concrete target for legal-system reform.
  • Any jurisdiction that publishes case records with court location, case type, filing date, and judgment date could apply the method, so the approach is not China-specific.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural stress test the paper does not run is to hold out entire provinces or years and check whether the reported R2 survives outside the observed four-year distribution.
  • In a deployed system, the column encoder cannot score one new city-year in isolation because it aggregates all rows in a column; an incremental update rule would be needed for real-time forecasts.
  • The paper reports predictive associations, so testing whether judicial performance actually causes FDI could use court staffing or procedural reforms as a natural experiment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes TLJD, a tabular learning method that predicts city-level foreign direct investment (FDI) in China from 380 judicial-performance indicators computed from over 12 million adjudication documents. The method encodes rows and columns of the tabular dataset, applies transformer layers with arithmetic attention, and uses a mixture-of-experts with four judicial-indicator-type experts. Cross-city and cross-time experiments compare TLJD with ten tabular baselines, reporting R2 values up to 0.92 on a mixed-year cross-city dataset and on a cross-time dataset.

Significance. If the results were valid, the paper would make two contributions: a large judicial-performance indicator dataset and evidence that court-document-derived indicators alone could predict city-level FDI more accurately than standard tabular methods. The dataset construction from 12 million documents is a substantial effort, and the idea of substituting judicial data for economic statistics is original. However, the reported performance figures are currently undermined by train/test leakage in the column encoder, so the paper's main empirical claim is not established. The method itself is plausible, and the underlying dataset could be valuable to the community if released.

major comments (4)
  1. [§4.1, Eq. (1); §5.1] The column encoder consumes the full column vector v_j in R^N with no statement that v_j is restricted to the training split. Section 5.1 only says that each dataset is divided into train/validation/test (3:1:1 for CCP; 2016–2018 vs 2019 for CTP), so both the min-max scaling parameters and the MLP inputs in Eq. (1) include test-row/test-year indicator values. Consequently, each training row's embedding h_i,j = phi^r_j(x_i,j) * phi^c(v_j) depends on test rows, so test information can influence the learned model during training. The high mixed-year CCP and CTP R2 values in Table 3 (0.9217 and 0.9242) may therefore reflect global column statistics and year effects rather than city-level judicial performance. As written, the model is transductive: scoring a single new city-year requires recomputing v_j over the entire dataset including the new row, which contradicts the forecasting scenario in §3.2.
  2. [§5.1, CTP task] For the cross-time task, the leakage is particularly damaging: the test set is all 2019 data, but the column encoder's v_j includes 2019 rows in the tabular dataset, so the model has access to future-year indicator values during training. Even if the min-max normalization were computed on training years only, the paper does not state this. The reported CTP R2 of 0.9242 therefore cannot be interpreted as evidence that judicial indicators from a city-year predict FDI for unseen future years.
  3. [§5.2, Table 4] The ablation study reports that removing the column encoder ('w/o ce') drops R2 from 0.9217 to 0.8821 on mixed-year CCP and from 0.9242 to 0.8464 on CTP. Because the column encoder is the sole component that uses cross-sample (and, in the current protocol, test-set) column statistics, this ablation may overstate the contribution of the column encoder. The authors should re-run the ablation with a training-only column encoder before claiming that global column similarities are responsible for the improvement.
  4. [§5.1, Tables 2–4] The paper reports each result from a single run without standard deviations, confidence intervals, or significance tests. Because the sample sizes are small (about 263–267 cities per year) and some advantages over baselines are marginal (e.g., Table 2, CCP 2016: TLJD R2=0.5471 vs SAINT 0.5432), the claim of 'best results in all evaluation metrics' is not established with the reported evidence. Multiple seeds and variance reporting are necessary, especially given the small dataset size.
minor comments (6)
  1. [Abstract, §5.2] The phrase 'reach to at least 0.92 R2' is misleading because the single-year CCP results in Table 2 are between 0.53 and 0.62; the 0.92 figure only holds for the mixed-year CCP and CTP, and should be qualified.
  2. [§5.1] There is a typo: 'division ration' should be 'division ratio'.
  3. [§4.1, Eq. (1)] The paper should clarify that MLP1(v_j) * MLP2(v_j) denotes scalar-vector multiplication, and should discuss the large parameter count of the column MLP: for N=1060, each column encoder has input dimension 1060, and with 380 columns the total column-encoder parameters are substantial; regularization for this component is not described.
  4. [§4.2] The notation for the CLS embedding in E^t_i is inconsistent: the matrix definition uses ecls_i, while Eq. (9) and Eq. (10) use eclst_i; please make the notation uniform.
  5. [§5.2] The expert-weight analysis in Figure 4(b) is based on the test set (2019) of CTP; if the leakage issue is corrected, this analysis should be redone on a properly split test set.
  6. [General] The paper does not include a data availability statement or a release of the constructed dataset and code; given the claimed contribution of the 380-indicator index system, the authors should provide the dataset or a detailed, reproducible description of the indicator computation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a standard supervised regression study whose claims rest on empirical benchmarks, not on a derivation that assumes its conclusion.

full rationale

TLJD is an empirical tabular-learning method. The outcome Y (city-level FDI) comes from China City Statistical Yearbooks, and the features X (380 judicial indicators) are computed from adjudication documents via a separately defined index system; no equation defines an indicator in terms of the target FDI, and no fitted parameter is renamed as a prediction. Hyperparameters, MoE weights, and column-encoder parameters are estimated on labeled data, which is ordinary supervised learning rather than circular reasoning. The paper contains no load-bearing self-citation: the cited prior work (AMFormer, ST-ExpertNet, etc.) is used for architectural components, and the central comparison is against ten external baselines on held-out splits. The transductive column-encoder concern raised near Eq. (1) - v_j is a vector over all N samples and Section 5.1 does not state that min-max scaling or column MLPs are restricted to training rows - is a potential data-leakage/soundness issue for the reported R2 values, not a circularity issue: even if true, it would mean the model peeks at test feature statistics, not that the prediction reduces to its input by definition. Accordingly the circularity score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim rests on a large hand-built indicator system and on an experimental protocol whose validity depends on an unstated train/test separation for column encoding. Hyperparameters are selected by grid search; the indicator system is not externally validated; target labels are official FDI statistics that the paper itself treats as potentially manipulable.

free parameters (4)
  • Embedding size d = 96
    Grid-searched on validation sets; affects model capacity and all reported results.
  • Number of transformer layers L = 2 (single-year CCP) or 3 (mixed-year CCP and CTP)
    Grid-searched hyperparameter; deeper stack improves mixed-year and CTP results.
  • Loss balancing weight lambda = 0.4 (single-year) or 0.6 (mixed/CTP)
    Grid-searched to balance regression and expert responsibility losses; changes expert weight behavior.
  • Training schedule (learning rate, batch size, epochs) = lr 1e-4 or 1e-3; batch 32; epochs 100 (single-year) or 50 (mixed/CTP)
    Chosen by grid search; standard training choices that affect convergence and final metrics.
assumptions (6)
  • domain assumption China Judgments Online provides a complete and representative record of local judicial activity
    All 380 indicators are derived from this corpus; selective publication or missing courts would bias city-level indicators. Section 3.1.
  • domain assumption The 380 hand-designed indicators measure judicial performance rather than confounding variables such as city economic scale
    Many indicators may be raw counts correlated with city size; no external validation against established judicial-quality measures is given. Section 3.1.
  • domain assumption FDI values from China City Statistical Yearbook are accurate ground truth
    The motivation is that official economic statistics can be manipulated, yet the target labels come from the same class of official statistics. Section 3.1.
  • ad hoc to paper Column vectors v_j can be computed without using test-set information
    Eq. (1) defines v_j over all N samples and the paper never specifies train-only column statistics; the experimental validity depends on this unstated assumption. Section 4.1.
  • domain assumption Random split without grouping by city is valid for mixed-year CCP
    The same city can appear in both training and test sets across years, which can leak persistent city-level effects into the test metric. Section 5.1.
  • ad hoc to paper Same-year judicial indicators are appropriate for 'future' FDI prediction
    CTP uses 2019 judicial features to predict 2019 FDI; this is not a temporal lead and does not demonstrate forecasting. Section 5.1.
invented entities (1)
  • Judicial performance index system (380 indicators)
    purpose: Quantify local judicial performance from adjudication documents for use as FDI predictors.
    The paper provides only a few example indicators; no external benchmark, inter-coder validation, or comparison with existing rule-of-law indices is provided. Figure 1(b).

how reviews work

0 comments
Cite this review

Pith. "Pith review of City-Level Foreign Direct Investment Prediction with Tabular Learning on Judicial Data." pith.science (2026). https://pith.science/paper/NX33GWBO

@misc{pith2026250705651,
  author       = {Pith},
  title        = {Pith review of: City-Level Foreign Direct Investment Prediction with Tabular Learning on Judicial Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NX33GWBO}},
  note         = {Machine review of arXiv:2507.05651}
}
read the original abstract

To advance the United Nations Sustainable Development Goal on promoting sustained, inclusive, and sustainable economic growth, foreign direct investment (FDI) plays a crucial role in catalyzing economic expansion and fostering innovation. Precise city-level FDI prediction is quite important for local government and is commonly studied based on economic data (e.g., GDP). However, such economic data could be prone to manipulation, making predictions less reliable. To address this issue, we try to leverage large-scale judicial data which reflects judicial performance influencing local investment security and returns, for city-level FDI prediction. Based on this, we first build an index system for the evaluation of judicial performance over twelve million publicly available adjudication documents according to which a tabular dataset is reformulated. We then propose a new Tabular Learning method on Judicial Data (TLJD) for city-level FDI prediction. TLJD integrates row data and column data in our built tabular dataset for judicial performance indicator encoding, and utilizes a mixture of experts model to adjust the weights of different indicators considering regional variations. To validate the effectiveness of TLJD, we design cross-city and cross-time tasks for city-level FDI predictions. Extensive experiments on both tasks demonstrate the superiority of TLJD (reach to at least 0.92 R2) over the other ten state-of-the-art baselines in different evaluation metrics.

Figures

Figures reproduced from arXiv: 2507.05651 by the authors.

Figure 1
Figure 1. The examples of (a) adjudication documents and (b) judi [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The framework of TLJD. (a) Indicator Feature Encoding maps each judicial performance indicator value to an embedding. (b) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Regional variations of expert weights. (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (a) Regional variations of FDI; (b) Average expert weights [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages

  1. [1]

    Drivers of economic and financial integration: A machine learning approach

    [Akbari et al., 2021] Amir Akbari, Lilian Ng, and Bruno Solnik. Drivers of economic and financial integration: A machine learning approach. Journal of Empirical Finance, 61:82–102,

    work page 2021

  2. [7]

    Deep learning-driven forecast- ing of Moroccan FDI: an LSTM-based approach

    [Chentouf et al., 2024] Amine Chentouf, Jihad Ait Sous- sane, and Zahra Mansouri. Deep learning-driven forecast- ing of Moroccan FDI: an LSTM-based approach. In Proc. of ICCSC, pages 1–6,

    work page 2024

  3. [8]

    Europe and central asia economic update, spring 2022: War in the re- gion

    [Demirguc-Kunt, 2022] Asli Demirguc-Kunt. Europe and central asia economic update, spring 2022: War in the re- gion. Technical report, World Bank Group,

    work page 2022

  4. [10]

    Revisiting deep learning models for tabular data

    [Gorishniy et al., 2021] Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep learning models for tabular data. In Proc. of NeurIPS , pages 18932–18943,

    work page 2021

  5. [11]

    Deep residual learning for image recog- nition

    [He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. In Proc. of CVPR, pages 770–778,

    work page 2016

  6. [12]

    [Huang et al., 2021] Juan Huang, Bifang Zhou, Huajun Huang, Jianjiang Liu, and Neal N. Xiong. An adap- tive lasso grey model for regional FDI statistics predic- tion. Computers, Materials and Continua , 69(2):2111– 2121,

    work page 2021

  7. [13]

    LightGBM: a highly efficient gradient boost- ing decision tree

    [Ke et al., 2017] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. LightGBM: a highly efficient gradient boost- ing decision tree. In Proc. of NeurIPS , page 3149–3157,

    work page 2017

  8. [18]

    CatBoost: unbiased boosting with categor- ical features

    [Prokhorenkova et al., 2018] Liudmila Prokhorenkova, Gleb Gusev, Aleksandr V orobev, Anna Veronika Dorogush, and Andrey Gulin. CatBoost: unbiased boosting with categor- ical features. In Proc. of NeurIPS, page 6639–6649,

    work page 2018

Show all 36 references
  1. [20]

    Prediction of foreign di- rect investment: An application of long short-term mem- ory

    [Roy, 2021] Sourabh Singha Roy. Prediction of foreign di- rect investment: An application of long short-term mem- ory. Psychology And Education, 58(2):4001–4015,

    2021

  2. [21]

    Trend prediction of FDI based on the intervention model and ARIMA-GARCH-M model

    [Shi et al., 2012] Hongyan Shi, Xin Zhang, Xiaoming Su, and Zhongju Chen. Trend prediction of FDI based on the intervention model and ARIMA-GARCH-M model. AASRI Procedia, 3:387–393,

    2012

  3. [23]

    Primer of applied regression and analysis of variance

    [Slinker and Glantz, 1990] Bryan K Slinker and Stan- ton Arnold Glantz. Primer of applied regression and analysis of variance. McGraw-Hill,

    1990

  4. [25]

    A survey on deep tabular learning

    [Somvanshi et al., 2024] Shriyank Somvanshi, Subasish Das, Syed Aaqib Javed, Gian Antariksa, and Ahmed Hossain. A survey on deep tabular learning. arXiv preprint arXiv:2410.12034,

    2024 arXiv

  5. [26]

    Autoint: Automatic feature interaction learning via self-attentive neural networks

    [Song et al., 2019] Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. Autoint: Automatic feature interaction learning via self-attentive neural networks. In Proc. of CIKM , page 1161–1170,

    2019

  6. [27]

    Modeling and fore- casting foreign direct investment into Brazil with ARIMA

    [Turolla and Margarido, 2011] FREDERICO ARAÚJO Tur- olla and Mário Antônio Margarido. Modeling and fore- casting foreign direct investment into Brazil with ARIMA. Economia Global e Gastao , 16:83 – 100, 09

    2011

  7. [30]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    [Vaswani et al., 2017] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of NeurIPS, page 6000–6010,

    2017

  8. [31]

    Determinants of judicial effi- ciency: a survey

    [V oigt, 2016] Stefan V oigt. Determinants of judicial effi- ciency: a survey. European Journal of Law and Eco- nomics, 42:183–208,

    2016

  9. [32]

    Unctad insights: Forecasting global FDI: A panel data approach

    [Vujanovi´c et al., 2021] Nina Vujanovi´c, Bruno Casella, and Richard Bolwijn. Unctad insights: Forecasting global FDI: A panel data approach. Transnational Corporations, 28(1):97–125,

    2021

  10. [33]

    ST- ExpertNet: A Deep Expert Framework for Traffic Predic- tion

    [Wang et al., 2023] Hongjun Wang, Jiyuan Chen, Zipei Fan, Zhiwen Zhang, Zekun Cai, and Xuan Song. ST- ExpertNet: A Deep Expert Framework for Traffic Predic- tion . IEEE Transactions on Knowledge & Data Engineer- ing, 35(07):7512–7525,

    2023

  11. [34]

    Global economic prospects, june

    [World Bank, 2024] World Bank. Global economic prospects, june

    2024

  12. [35]

    A closer look at deep learning on tabular data.arXiv preprint arXiv:2407.00956,

    [Ye et al., 2024] Han-Jia Ye, Si-Yang Liu, Hao-Run Cai, Qi- Le Zhou, and De-Chuan Zhan. A closer look at deep learning on tabular data.arXiv preprint arXiv:2407.00956,

    2024

  13. [36]

    Modeling and prediction using an artificial neural network to study the impact of foreign direct investment on the growth rate/a case study of the state of qatar

    [Zain Al-Thalabi et al., 2022] Sahera Hussein Zain Al- Thalabi, Ali Akbar Heydari, and Masoud Tavakoli. Modeling and prediction using an artificial neural network to study the impact of foreign direct investment on the growth rate/a case study of the state of qatar. Journal of...

    2022

  14. [1987]

    Transforming our world: The 2030 agenda for sustainable development

    [United Nations, 2015] United Nations. Transforming our world: The 2030 agenda for sustainable development. Technical report, United Nations,

    2015

  15. [1990]

    Saint: Improved neural networks for tabular data via row attention and contrastive pre-training

    [Somepalli et al., 2021] Gowthami Somepalli, Micah Gold- blum, Avi Schwarzschild, C Bayan Bruss, and Tom Gold- stein. Saint: Improved neural networks for tabular data via row attention and contrastive pre-training. arXiv preprint arXiv:2106.01342,

    2021 arXiv

  16. [2000]

    Layer normalization

    [Lei Ba et al., 2016] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450,

    2016 arXiv

  17. [2001]

    Governments manipulate offi- cial statistics: Institutions matter

    [Briviba et al., 2024] Andre Briviba, Bruno Frey, Louis Moser, and Sandro Bieri. Governments manipulate offi- cial statistics: Institutions matter. European Journal of Political Economy, 82:102523,

    2024

  18. [2011]

    [Tyler, 1987] T. Tyler. Procedural justice research. Social Justice Research, 1:41–65,

    1987

  19. [2012]

    Comparison between artificial neural network and linear model prediction performance for FDI disparity and the growth rate of companies in hun- garian counties

    [Singh, 2023] Devesh Singh. Comparison between artificial neural network and linear model prediction performance for FDI disparity and the growth rate of companies in hun- garian counties. International Journal of Business Infor- mation Systems, 43(4):542–552,

    2023

  20. [2015]

    Understanding pro- cedural justice and its impact on business organizations

    [Konovsky, 2000] Mary A Konovsky. Understanding pro- cedural justice and its impact on business organizations. Journal of Management, 26(3):489–511,

    2000

  21. [2016]

    Arithmetic feature inter- action is necessary for deep tabular learning

    [Cheng et al., 2024] Yi Cheng, Renjun Hu, Haochao Ying, Xing Shi, Jian Wu, and Wei Lin. Arithmetic feature inter- action is necessary for deep tabular learning. In Proc. of AAAI, pages 11516–11524,

    2024

  22. [2017]

    Papadopoulos, 2023] Irene Kilanioti and George A

    [Kilanioti and A. Papadopoulos, 2023] Irene Kilanioti and George A. Papadopoulos. A knowledge graph-based deep learning framework for efficient content similarity search of sustainable development goals data. Data Intelligence, 5(3):663–684,

    2023

  23. [2018]

    Modelling and forecasting foreign di- rect investment: A comparative application of machine learning based evolutionary algorithms hybrid models

    [Rapoo et al., 2023] Mogari Ishmael Rapoo, Martin Chanza, and Elias Munapo. Modelling and forecasting foreign di- rect investment: A comparative application of machine learning based evolutionary algorithms hybrid models. In Proc. of ICO, pages 23–35,

    2023

  24. [2019]

    XGBoost: A scalable tree boosting system

    [Chen and Guestrin, 2016] Tianqi Chen and Carlos Guestrin. XGBoost: A scalable tree boosting system. In Proc. of KDD, page 785–794,

    2016

  25. [2021]

    Random forests

    [Breiman, 2001] Leo Breiman. Random forests. Machine Learning, 45(1):5–32,

    2001

  26. [2022]

    Friedman

    [Friedman, 2001] Jerome H. Friedman. Greedy function ap- proximation: A gradient boosting machine. The Annals of Statistics, 29(5):1189–1232,

    2001

  27. [2023]

    Kingma and Jimmy Ba

    [Kingma and Ba, 2015] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of ICLR,

    2015

  28. [2024]

    [Bruneckiene et al., 2019] Jurgita Bruneckiene, Robertas Jucevicius, Ineta Zykiene, Jonas Rapsikevicius, and Man- tas Lukauskas. Assessment of investment attractiveness in european countries by artificial neural networks: What competences are needed to make a decision on colle...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.