A Multimodal Deep Learning Framework for Default Prediction in Peer-to-Peer Lending

Credit access is essential to economic development and enables individuals and companies to invest, mitigate risk, and build financial strength. However, many borrowers, especially students, freelancers, and small-business owners, remain underserved by the traditional lending sector. Conventional banks rely strongly on structured indicators such as credit scores, income history, and debt-to-income ratios, making it harder for individuals with limited or non-traditional financial histories to qualify. Peer-to-peer (P2P) lending platforms provide a more open and accessible alternative by connecting borrowers and lenders directly through digital platforms [7]. Predicting loan default nevertheless remains challenging. The initial credit risk models were mostly of the statistical type: logistic regression, discriminant analysis and probit models [3]. These techniques were simple and easily interpretable, but had limited capacity to identify non-linear relationships and complex interactions between features. Due to the power of computing and the growing number of large-scale data sets, machine learning methods such as decision trees, random forests, and support vector machines (SVMs) are being increasingly used to increase the accuracy of prediction. While such progress has been made, the models that are mostly built on structured financial and demographic information still have a significant limitation in representing the qualitative, behavioral and contextual characteristics of borrowers. Figure 1 shows the simplified workflow of P2P lending systems. Current credit-scoring strategies focus heavily on financial attributes and often disregard what borrowers say when describing their loans. These descriptions may contain credit-relevant signals regarding borrower intent, financial discipline, and financial activity, but they are difficult to analyse using traditional statistical methods [8]. To address this limitation, this study proposes an AI-enabled P2P lending system that incorporates structured financial features and unstructured text for default prediction. Five deep learning models, Average Embedding (AE), CNN, Bidirectional GRU (BiGRU), CNN–RNN Hybrid, and DistilBERT, are evaluated using ROC-AUC and other classification metrics. The most accurate text model is then fused with structured features through multimodal late fusion. The proposed pipeline is implemented as an end-to-end solution, featuring a Flutter mobile front end for borrower-data input and a FastAPI backend for real-time default-risk inference.

Credit default prediction is a primary problem in financial risk management. Conventional models use structured borrower characteristics such as income, employment status, credit score, and repayment history. Logistic regression and probit models were among the earliest interpretable statistical approaches, while more recent machine learning models have improved predictive power. However, structured information can omit behavioural evidence [10]. Researchers have therefore incorporated soft information from unstructured text. In P2P lending, borrower-written descriptions can reveal intent, credibility, and repayment attitude. Liu et al. [9] provided a theoretical basis for this idea, while Jiang et al. [6] showed empirically that combining text with structured features improves default prediction. As digital technology has advanced, there’s been a flood of textual information on online lending platforms. When people take out a loan, they tend to give descriptions of why they need the loan. Traditional credit scoring models do not make use of this kind of unstructured data, but recent studies indicate that borrowerprovided textual descriptions can provide useful insights into borrowers’ reliability, intentions and creditworthiness. The extraction of meaningful information in unstructured text is difficult since such information cannot simply be fed into conventional statistical models. Hence, for processing these text-based descriptions and extracting relevant information for credit assessment, advanced NLP and ML technologies are needed. Early text-based approaches used handcrafted features including sentiment, spelling quality, and text length [5]. These methods identified useful signals but captured limited context. Traditional NLP representations such as TF-IDF and LIWC produced moderate gains [12, 16], but remained frequency-based. Deep learning enabled automatic representation learning. CNNs [10], RNNs [1], and hybrid models [11] have performed well in financial NLP, although much of this work concerns longer documents rather than short borrower narratives. There are several limitations in existing literature to show that there is improvement in the credit default prediction with textual information. Numerous studies are based on lengthy and formal loan descriptions, such as annual reports and corporate disclosures, while not on short loan descriptions created by the borrower. Typical studies that investigate the narratives of P2P lending tend to use either TF-IDF representations or manually designed features or relatively shallow machine learning models. Likewise, deep learning methods have often been used for loan officer notes or formal financial text, instead of for the informal description that is often provided directly from the borrower. In addition, there has been little effort to systematically assess the impact of the text-derived probability in the context of the various feature-fusion settings used: text-derived probability combined with FICO score, structured borrower information without FICO score, and the full set of structured information. These gaps encourage the need for a systematic assessment of the text information generated by the borrowers and its complementary value in the context of the structured credit-related features. Transformer models such as BERT further improved contextual modelling [14]. For short P2P borrower text, however, simpler models may match more complex transformers, which is relevant to efficient deployment [8]. Recent work has also examined generative AI for improving financial text [15], largely using officer-drafted notes rather than borrower-authored descriptions. A systematic comparison of multiple text architectures, combined with structured features under a common protocol and demonstrated in a deployable system, therefore remains useful.

Traditional credit scoring depends mainly on structured financial information, while borrower-written loan descriptions are often ignored despite their potential credit signals [8]. This work proposes a multimodal framework combining borrower-written descriptions with structured financial features. The framework evaluates lightweight and transformer-based text models, improves prediction through late fusion, and is deployed as a Flutter–FastAPI application for real-time risk assessment.

This section presents the dataset, preprocessing pipeline, model architectures, and system design. Figure 2 illustrates the end-to-end architecture.

Following[8], we use the “All Lending Club Loan Data” dataset from Kaggle, comprising 2,260,701 loans issued between 2007 and 2018 with 151 features. Four sequential filters reduce the raw data to a model-ready sample, as shown in Table 2: a temporal filter retaining 2007–2014 records; a loan-status filter restricting the target to fully paid (0) or charged-off/defaulted (1); and text availability and length filters requiring at least 40 cleaned words per description. The final sample contains 44,902 records with a default rate of 14.51%. The 19 raw structured columns are transformed into 70 engineered features: nine direct numerical features, four derived numerical features, three binary indicators, and 54 one-hot encoded categorical features. Each loan description is cleaned by removing HTML tags, punctuation, and extra whitespace; converting currency symbols to words; and lowercasing text. Sequences are padded or truncated to 115 tokens. The dataset is split into training (20,000), validation (10,000), and test (14,902) sets using stratified sampling with seed 42.

Experiments are organised into two phases. Phase 1 independently assesses text-based deep learning models and three structured feature subsets, No-FICO, FICO-only, and All-Structured. Phase 2 combines each text probability with each structured subset using Logistic Regression.

Logistic Regression with an L2 penalty, the lbfgs solver, balanced class weights, and standardised inputs is trained on FICO-only, Structured-No-FICO, and AllStructured feature subsets. The regularisation strength𝐶 is selected by validation ROC-AUC over ten values in [0.05,0.50]. As 𝐶 had no measurable impact,𝐶 = 0.05 is used throughout. All five text models follow an identical protocol. Stage A selects candidate configurations using validation ROC-AUC. Stage B evaluates no adjustment, balanced class weights, and randomundersampling at 1.5 times the minority count. Stage C scans the decision threshold over 𝑡 ∈ [0.10,0.50] on the validation set, prioritising recall. The test set is used only for final evaluation. Average Embedding maps tokens to trainable embeddings and averages all 115 token vectors before a sigmoid output layer. CNN uses Conv1D and max pooling to capture local patterns. BiGRU uses SpatialDropout1D (0.3), a bidirectional GRU with 128 units, a 64-unit ReLU layer, dropout (0.3), and a sigmoid output. The CNNRNN Hybrid applies Conv1D and max pooling before a 64-unit GRU. DistilBERT [13] uses pretrained distilbert-base-uncased weights and a two-class linear head. Its search uses Optuna TPE [2], with 20 trials per seed and seeds 42, 123, and 7.

Each text model produces a scalar default probability 𝑝text ∈ [0,1]. This value is appended to each structured feature subset, and a Logistic Regression meta-classifier is fitted on the validation split. A common pipeline covering feature construction, hyperparameter tuning, fitting, threshold calibration, and test evaluation is applied to all configurations.

The system is designed to be a borrower-to-lender system, with a mobile frontend implemented using the Flutter framework, and a backend using the FastAPI framework. Users sign up in a single authentication process including a splash screen, sign-in/sign-up, and OTP verification, with a choice of role as borrower or lender. Personal, financial and loan request data are entered into validated forms, and sent to the backend as JSON payloads via HTTP POST request. The informationsubmittedisprocessedbytwocomplementary pipelines in the backend. The loan description is first tokenized, then padded to a sequence length of 115 and fed to a text-based text classification model using Keras to deliver a text-based default probability. The probability is then added to the form features provided by the borrower to form a 71-feature vector and classified with a Logistic Regression model loaded from a serialized scikit-learn pipeline. The risks core and loan grade, along with the estimated default probability, are then sent back to the mobile application and displayed on the AI Risk Score screen. The assessment will help the borrower to either proceed with the listing of the loan or adjust the application and re-apply. Through their dashboard, lender scan view loan applications that are available, view borrowers’ financial profiles and AI risk assessments, fund loans, and track repayment activity via tracking and history interfaces.

Table-5 shows that FICO alone is a weak predictor, with a test AUC of 0.598, whereas the structured feature sets reach approximately 0.684–0.685. Average Embedding performs best among the text-only models with a test ROC-AUC of 0.615, closely followed by DistilBERT at 0.612. The CNN–RNN Hybrid’s low recall at its calibrated threshold illustrates the difficulty of imbalanced classification. The AE+All Structured and AE+NoFICO configurations both achieve a test ROC-AUC of 0.713, improving on the All-Structured baseline of 0.684 and the best text-only result of 0.615. The matching result without FICO suggests that borrower descriptions may partly compensate for the absence of this feature.

Average Embedding achieved the strongest text-only performance, reaching a test ROC-AUC of 0.615, slightly higher than DistilBERT at 0.612 and outperforming CNN, BiGRU, and CNN–RNN Hybrid. This is consistent with Kriebel et al.[8] and suggests that simple embedding-based representations are well suited to short, informal loan descriptions. Larger models may be more prone to overfitting on brief narratives, while DistilBERT’s general-domain pretraining may transfer imperfectly to sparse financial text. TheCNN–RNN Hybrid also struggled under the 14.51% default rate. Lowering the decision threshold improved recall but increased false positives, highlighting the need for threshold calibration. Combining Average Embedding predictions with structured features increased ROC-AUC to 0.713, exceeding the tabular-only result of 0.684. Comparable performance with out FICO suggests that borrower written descriptions may partially compensate for unavailable credit-bureau information. Nevertheless, the dataset covers loans issued between 2007 and 2014 and contains 44,902 records, which may limit generalisation to current lending environments. Similar wording across default and non-default loans also reduces class separability. Future work could investigate attention-based fusion or non-linear meta-classifiers such as CatBoost.

This paper proposes an AI-based P2P lending model combining structured financial data with borrower-authored loan descriptions. Five text models and three structured feature subsets yield 15 multimodal configurations. The lightweight Average Embedding model achieved the best text-only ROC-AUC of 0.615, matching or exceeding more complex methods. Combining text and structured features increased ROC-AUC to 0.713, with threshold calibration proving essential under class imbalance. These findings indicate that loan narratives can add predictive information and may partly compensate for missing credit-score information. Future research should evaluate larger and geographically diverse datasets and investigate stronger non-linear fusion techniques and language-model encoders.