| Journal of Information and Communications Technology:
Algorithms, Systems and Applications
Received: 10 May 2026; Revised: 17 August 2026; Accepted: 24 August 2026; Published Online: 25 August 2026.
J. Inf. Commun. Technol. Algorithms Syst. Appl., 2026, 2(3), 26312 | Volume 2 Issue 3 (September 2026) | DOI: https://doi.org/10.64189/ict.26312
© The Author(s) 2026
This article is licensed under Creative Commons Attribution NonCommercial 4.0 International (CC-BY-NC 4.0).
Mohammad Ali Ansari,
*
Huzaifa Zahid Husein Shah, Talha Ansari, Aditya Vijay Kamble and Ahlam Ansari
Department of Computer Engineering, M. H. Saboo Siddik College of Engineering, Byculla, Mumbai, Maharashtra, 400008, India
*Email: ali.221204.co@mhssce.ac.in (Mohammad Ali Ansari)
Abstract
Many investors in India face difficulties in choosing a suitable investment avenue among diverse options. This
occurs because of fragmented financial platforms, diverse asset behaviors, and limited access to intelligent
decision support tools. Existing research focuses mainly on single-asset forecasting, leaving a gap for a unified,
multi-asset investment guidance system. This research proposes an AI-supported unified investment guidance
platform that integrates all major asset classes, such as stock, metals, real estate, mutual funds, and
cryptocurrency, into a single user-friendly framework. This platform utilizes historical and real-time data
collected through APIs, web scraping, and public datasets combined with data analysis, machine learning and
deep learning models for accurate presentation and forecasting. Random forest is employed for structured
datasets such as gold and real estate, while Long Short-Term Memory (LSTM) networks are used for time-series
prediction of stocks, cryptocurrency and mutual fund NAVs. To enhance contextual understanding, a sentiment
analysis module processes financial news and market headlines, and an AI-based advisory layer converts
analytical outputs into natural language investment insights. Experimental research exhibits strong predictive
performance, achieving R² values of 0.92 for gold, 0.88 for real estate, 0.78 for stocks, 0.80 for cryptocurrency,
and 0.99 for mutual funds. The platform combines real-time and historical data with quantitative forecasting,
sentiment analysis, and an AI-driven advisory layer to provide interpretable decision-support information. The
proposed platform contributes to democratizing financial guidance in India by enabling personalized, data-
driven, and holistic investment decision-making.
Keywords: Finance; Multi assets; Wealth management; Investment; Investment guidance.
1. Introduction
1.1 Background and importance of investment decision-making
Over the past decade, the growth of global financial markets has dramatically increased the number of
investment opportunities available to individual investors, and while this trend is most evident in developed
countries, it also represents increasing complexity for individual investors in developing economies such as
India. Today, individual investors are faced with a multiplicity of asset classes, including stocks, mutual funds,
gold, real estate, and cryptocurrencies, each with its own unique characteristics of both risk and return. Sound
investment decision-making requires investors to continuously interpret this diverse and fast-changing
information across asset classes, a task that directly influences long-term wealth creation, retirement planning,
and financial security for households.
[1]
As financial markets in India continue to deepen and diversify, the ability
to make timely, well-informed investment decisions has become an increasingly important determinant of
individual financial well-being.
1.2 Current challenges in investment
Traditional approaches to investing rely heavily on historical data, manual review, or platform-specific analysis
that does not take into account all of the key asset classes at once, therefore rendering these traditional
approaches ineffective at reflecting changing market conditions and the diverse preferences of individual
investors. Individual investors in India face additional practical obstacles: financial platforms remain
fragmented across asset classes, with dedicated applications for stocks, mutual funds, real estate, and
cryptocurrency operating largely in isolation from one another.
[2-8]
Each asset class also exhibits distinct
behavioral and statistical characteristics, so models developed for one asset class, such as real estate
[2-4,9]
or
gold,
[5-7]
rarely transfer well to another. Compounding this fragmentation, a large proportion of Indian retail
investors have limited financial literacy and limited access to professional advisory services, leaving them
without the intelligent decision-support tools needed to compare opportunities across the full range of available
asset classes.
1.3 Role of artificial intelligence in financial services
The advent of artificial intelligence (AI), machine learning (ML), and other related technologies provides a
promising opportunity to improve the quality of investment analysis through data-driven forecasting,
automated decision support, and personalized recommendations. Prior work has shown that ML and deep
learning models can outperform traditional statistical approaches for price and return prediction across
individual asset classes, including regression and ensemble methods for real estate and gold,
[2-4,9]
deep learning
and hybrid architectures for time series forecasting,
[5-7,9]
sentiment-aware models that incorporate news and
social media signals,
[8,10]
and reinforcement-learning-based systems for portfolio optimization and personal
finance management.
[11,12]
While there are numerous models being utilized in financial applications that show
great promise, very few models are currently used in practice. This limited use is attributed to the fact that there
are no integrated platforms that allow users to access multiple asset classes in real time. The current state of
fragmentation in the Indian financial technology landscape, combined with low levels of financial literacy,
highlights the need for a comprehensive, user-centric investment guidance system.
1.4 Research gap
A closer look at the literature reveals that most AI-driven financial forecasting research is confined to a single
asset class: real estate,
[2-4,9]
gold,
[5-7]
or cryptocurrency,
[10,11]
with relatively few studies addressing multiple asset
classes within one unified framework.
[1]
Even when multiple-asset or portfolio-level approaches exist,
[1,12]
they
tend to sacrifice interpretability for predictive performance, rely on narrow sentiment sources such as a single
social media platform,
[10]
or remain largely conceptual without deployment or validation against real, live
markets.
[13]
A detailed discussion of these gaps, organized by modeling approach, is presented in Section 2.6. In
summary, no existing system combines multi-asset price forecasting, news-based sentiment analysis, and
natural-language advisory generation within a single, deployed, user-facing platform tailored to the Indian
market.
1.5 Proposed methodology
To address this gap, this research proposes an AI-supported unified investment guidance platform that
integrates five major asset classes-stocks, mutual funds, gold and metals, real estate, and cryptocurrency-into a
single, user-friendly framework. The platform combines historical and real-time data collected through API,
web scraping, and public datasets with a multimodel machine learning and deep learning pipeline: Random
Forest regression is applied to structured, slower-moving datasets such as gold and real estate, whereas LSTM
networks are used for the more volatile, sequentially dependent time series of stocks, cryptocurrency, and
mutual fund NAVs. A FinBERT-based sentiment analysis module processes financial news headlines to capture
market mood, and a large language model (LLM)-based advisory layer converts the combined analytical and
sentiment outputs into personalized, natural-language investment guidance. The full methodology is described
in Section 3.
1.6 Key contributions
The key contributions of this research are as follows:
•
Integration of five major asset classes-stocks, mutual funds, gold/metals, real estate, and cryptocurrency-
within a single analytical and advisory platform, rather than the single- or dual-asset focus of prior work
. [2-10,13]
•
A domain-tailored multi-model forecasting architecture that pairs a random forest model with structured,
low-frequency asset data (gold, real estate) and LSTM networks with high-frequency, sequential asset data
(stocks, cryptocurrency, mutual funds).
•
A FinBERT-based sentiment analysis module that draws on diverse financial news sources rather than a
single social media platform, addressing a limitation identified in prior sentiment-based forecasting research.
[10]
•
An LLM-based advisory layer built on the NVIDIA Nemotron model that converts quantitative forecasts and
sentiment signals into interpretable, natural-language investment guidance, addressing the performance–
interpretability trade-off noted in prior deep learning studies.
[1]
•
A working, deployed platform validated against baseline investment strategies (Naïve Forecasting and
Simple Moving Average), rather than a purely theoretical model, addressing the deployment gap identified in
the literature.
[11]
1.7 Paper organization
The remainder of this paper is organized as follows. The rest of Section 2 presents a detailed review of the
literature across five methodological categories, identifies the research gap in greater depth, and differentiates
the proposed platform from existing investment applications. Section 3 describes the methodology, including
the data collection, model design, sentiment analysis module, AI advisory layer, data visualization, and system
architecture. Section 4 presents the evaluation metrics and experimental results for each asset class and
concludes with a discussion of their implications in Section 5. Section 6 concludes the paper and outlines
directions for future work.
2. Literature review
The application of artificial intelligence (AI) and machine learning (ML) in financial forecasting has developed
into an area that is among the most active and popular in computational finance. A brief review of some of the
major contributions for gold, real estate, stocks and cryptocurrencies is presented below.
2.1 Traditional and regression-based models
Traditional and regression-based approaches, including linear regression, multiple regression, and tree-based
regressors such as XGBoost and Random Forest -form the baseline methodology for structured, tabular financial
datasets such as year-by-year real estate prices and daily gold prices. These models typically use price history
together with location-based or macroeconomic features as predictors and are typically evaluated using the
coefficient of determination (R²) alongside error-based metrics such as the RMSE and MAE. A summary of the
traditional and regression-based machine learning methods applied to the prediction of gold and real estate
prices can be seen in Table 1.
Table 1: Traditional and regression-based models
Authors (Ref.)
Asset
Key Findings
Limitations
Maloku et al.
[2]
Real
Estate
They found that random forest algorithms
could be bet- ter at predicting real estate
prices then linear regression algorithms
when they used location-based features
Their sample size was limited
and their approach is static.
Srikanth and Ramchander
[3]
Real
Estate
Srikanth and Ramchander’s use of xgboost
to predict real estate prices had an R² of
0.934 which is quite high
They did not incorporate any
dynamic macroeconomic
indicators into their model.
Satyam and Suresh
[4]
Real
Estate
Satyam and Suresh were able to
demonstrate very good real-time
prediction capabilities in terms of
predicting real estate prices (R² = 0.9012)
Their dataset was static and did
not contain as many con- textual
indicators as it could have.
2.2 Ensemble and hybrid techniques
Ensemble and hybrid techniques combine multiple base learners-for example, stacking regression models with
boosting algorithms or pairing deep neural networks with traditional machine learning models-to reduce
forecasting error and capture relationships that a single algorithm may miss. These approaches are typically
applied to real estate, multi-asset portfolios, and broader financial market datasets and are evaluated using the
same error-based metrics (RMSE, MAE) and R² used for single-model approaches, alongside portfolio-level
metrics such as risk-adjusted return where applicable. Table 2 shows some examples of different ensemble
methods and hybrid approaches that can be used for financial forecasting.
Table 2: Ensemble and hybrid techniques
Authors (Ref.)
Asset
Key Findings
Limitations
Deepu et al.
[9]
Real Estate
By combining multiple models together,
Deepu et al. were able to reduce errors in
their forecasting
This approach does increase the
complexity of the overall model.
Mahajan
[1]
Multi-Asset
Mahajan’s results indicated that deep
neural network models generated
returns that were significantly greater
than those of tra- ditional machine
learning models when applied to multii-
asset portfolios
This approach reduces the
interpretability of these models.
Jay et al.
[14]
Cryptocurrency
Jay et al. applied a stochastic neural
network architecture to model the
inherent randomness of cryptocurrency
price movements, improving prediction
accuracy over deterministic neural
network baselines.
Manual hyperparameter tuning
of the model can be inefficient.
2.3 Deep learning and time series forecasting
Deep learning approaches, particularly LSTM networks and their bidirectional variant (BiLSTM), use gated
recurrent architectures to capture long-range temporal dependencies in sequential financial data. These models
are trained on historical price sequences-typically using a fixed look-back window-for assets such as gold, silver,
and cryptocurrency and are evaluated using the RMSE, MAE, and R² computed on a held-out test period. Table
3 provides a detailed view of how deep learning methods are being used to forecast time series data using
LSTMs and BiLSTMs.
Table 3: Deep learning and time series forecasting
Authors (Ref.)
Asset
Key Findings
Limitations
Zangana & Obeyd
[5]
Multi-asset
Zangana & Obeyd found
that Bi-LSTM models were more
effective than LSTM models at
forecasting multivariate data with an
R² of 0.95
They did not explore other
external indicators in addition to
historical data.
Amini & Kalantari
[6]
Gold & Silver
Amini & Kalantari found
that they were able to capture the
extreme movements in the long-term
data of gold and silver prices using
their deep learning methodology
The methodology they
developed is computation- ally
expensive.
Nagata et al.
[7]
Gold
Nagata et al. found that
an approach that combined both the
Gold and Daily data provided the best
forecast of future stock prices
Did not fully utilize either
the volume or macro factors.
Tanwar et al.
[11]
Crypto
(Altcoins)
Tanwar et al. used a
technique known as dependency to
improve their ability to predict the
price of Altcoin cryptocurrencies based
on the price of Bitcoin.
It had model complexity
and scalability issues
2.4 Sentiment analysis and external data integration
Sentiment analysis approaches extract investor mood from unstructured text-typically social media posts or
news headlines-using either lexicon-based scoring or pretrained language models and combine the resulting
sentiment score with historical price data as an additional predictive feature. These hybrid sentiment-price
models are typically applied to cryptocurrency and real estate stock forecasting and are evaluated by comparing
prediction accuracy (e.g., Root Mean Square Error (RMSE) and directional accuracy) against an equivalent
model trained on price data alone. Table 4 presents a summary of the literature that examines the use of
sentiment analysis for predicting stock prices.
Table 4: Sentiment analysis and external data integration
Authors (Ref.)
Asset
Key findings
Limitations
Arslan
[10]
Crypto
His study demonstrated that incorporating
sentiment from social media posts resulted in
a significant improvement in his model’s
ability to predict cryptocurrency prices
relative to a model that only utilized historical
price data.
The source of the sentiment
information in this study was limited to
Twitter and only utilized data from a
relatively short period of time.
Fan & Chen
[8]
Real
Estate
The authors reported that their model
resulted in higher prediction accuracy than a
model that did not utilize sentiment.
The model may not be applicable to
other markets because it was trained
using data collected in the Chinese
stock market.
2.5 Reinforcement learning and practical applications
Reinforcement learning (RL) approaches frame investment decision-making as a sequential decision problem
in which an agent learns an asset-allocation or budgeting policy by maximizing a reward signal-typically
portfolio return-through repeated interaction with historical or simulated market data. These approaches are
typically evaluated by benchmarking backtested portfolio returns against a market index or a rule-based
baseline. Table 5 lists some examples of how reinforcement learning can be applied to develop AI systems that
support decision-making in the area of personal finance.
The literature indicates a movement toward the use of ensemble and deep learning techniques for modeling
and prediction. In terms of India’s investment landscape, there appears to be a common theme among the
literature regarding the need for a unified platform for providing investment advice by combining the various
types of assets into a single interface, using external economic indicators and balancing predictive accuracy with
interpretability.
Table 5: Reinforcement learning and practical AI applications
Authors (Ref.)
Domain
Key contributions
Limitations
Leung et al.
[12]
Portfolio Opt.
Leung et al. developed a model that
utilized reinforcement learning to
deter- mine optimal asset allocation
portfolios. Their model produced
returns that exceeded those of the
S&P 500 index during backtesting.
They also noted that liquidity
constraints and long-term market risks
are not well represented by the model.
Deepthi et al.
[13]
Personal Finance
Deepthi et al. developed a system that
utilized reinforcement learning to
assist individuals with budgeting and
managing their portfolios.
The system remains largely conceptual
and lacks empirical evidence that
demonstrates its effectiveness in real-
world markets.
2.6 Identified research gap
Several research areas have been identified as gaps on the basis of the collective findings of the above
literature:
•
Limited datasets - Most studies use relatively small, static datasets that do not accurately reflect the
complexities of real-world situations.
•
Omission of key external factors-Macroeconomic trends, liquidity and trading volume-is typically omitted
from models when determining how they may impact investment decisions.
•
Oversimplification of models - While some models are multivariate and consider relationships between
different asset classes, many remain univariate and do not account for the interplay between related assets.
•
Limitations of sentiment analysis - Sentiment analysis models are limited because of their reliance on narrow
sources (e.g., Twitter), which limits their ability to provide robust sentiment-based models.
•
Inefficiencies in optimization - Hyperparameters are manually tuned, and simplifying assumptions are made
to optimize models, limiting their overall accuracy.
•
Trade-offs between performance and interpretability - Highly complex models provide very little
transparency and therefore diminish investors’ confidence in using the results in real-world applications.
•
Gaps in deployment - many models developed exist only in theory; few have been deployed into live markets
to validate their effectiveness.
•
Focus on short-term forecasts - Most models focus on predicting investment outcomes over the short-term
horizon, making them less relevant to investors who make investment decisions based on longer-term
strategies.
•
Validation of profitability - Few studies evaluate the practical profitability of investment portfolios and/or
investor outcomes on the basis of the investment recommendations generated by the model.
2.7 Differentiation from existing platforms
Popular investment analysis platforms do not include all major asset classes; most of them focus on only 1–3
asset classes, such as Groww (Stocks, Mutual Funds, Metals), MagicBricks (Real Estate), and Binance (Crypto).
Additionally, they do not combine technologies such as market sentiment analysis, an AI advisory section, future
price prediction, or cross-asset comparison within a single platform.
The proposed system directly addresses the abovementioned gaps. The proposed solution integrates five major
asset classes-stocks, mutual funds, metals, real estate, and cryptocurrency-under one analytical roof. It employs
different ML-based future prediction models (LSTM and random forest), incorporates a sentiment analysis
module drawn from diverse financial news sources beyond social media alone, and includes an additional LLM-
based advisory layer for personalized investment guidance.
2.8 Scope of the study
This research investigates the development and implementation of an artificial intelligence (AI)-driven unified
investment guidance platform that can support the needs of the Indian market. This proposed system will
incorporate multiple asset classes, including but not limited to stocks, mutual funds, real estate,
cryptocurrencies and metals, within a single analytical structure. The system uses supervised machine learning
models to develop classification models, deep learning-based time series models to predict future price
movements, and both financial news sentiment analysis and a large language model–based advisory layer to
generate personalized, interpretable and data-driven investment recommendations. The platform focuses on
generating medium- to long-term forecasts and providing individual and comparative assessments of asset
classes rather than facilitating high-frequency trading.
3. Methodology
3.1 Data collection
This dataset is an aggregation of five major asset types, namely, Stocks, Mutual Funds, Real Estate,
Cryptocurrency and Metals. Additionally, they are based on multiple time frames and provide a foundation for
analyzing investments.
Stock price data were obtained from Alpha Vantage's API and included real-time stock prices gathered through
web scraping on Google Finance using Selenium and Beautiful Soup. Cryptocurrency data from the past 10 years
were retrieved from the CoinGecko API. NAV data for mutual funds were collected from the Association of
Mutual Funds in India (AMFI), utilizing the mftool Python library.
Data concerning real estate (from 2015–2025) were collected from MagicBricks; specifically, the costs of real
estate and yields were based upon localities. Historical past 10 years of gold price data were collected from
Kaggle, while real-time gold and silver prices were collected from The Economic Times. Historical data
concerning the price of silver were collected from SilverPrice.org. Table 6 summarizes the data type, source, and
coverage period used for each asset class.
Table 6: Data sources and types by asset class
Asset class
Data type
Source
Coverage/Frequency
Stocks
Real-time and historical
OHLCV prices
Alpha Vantage API; Google
Finance (via Selenium and
Beautiful Soup web scraping)
Real-time, daily
Cryptocurrency
Historical daily price data
CoinGecko API
Past 10 years (2015–2025)
Mutual Funds
Net Asset Value (NAV)
Association of Mutual Funds
in India (AMFI), via the mftool
Python library
Daily NAV history
Real Estate
Locality-wise average price
per sq. ft.
MagicBricks
2015–2025 (yearly)
Gold
Historical daily closing price
Kaggle
Past 10 years (2015–2025)
Gold and Silver
Real-time spot price
The Economic Times
Real-time
Silver
Historical price data
SilverPrice.org
Historical
3.2 Data splitting
The train/test split strategy was tailored to each asset class's dataset frequency and structure rather than a
single split applied uniformly across all the models. For assets with daily historical –data-gold and mutual fund
–NAVs-training and test sets were split chronologically, with approximately 80% of each dataset used for
training and the remaining 20% (all records after a fixed cutoff date) reserved for testing, to preserve temporal
order and prevent data leakage. For stock price prediction, the model was trained on historical data from a
representative subset of NSE-listed companies and evaluated out-of-sample on AAKASH. NS, a company not
included in training, over a fixed forward test window (see Section 4.2). For cryptocurrency price prediction,
the model was trained on historical data up to a fixed date and evaluated over a distinct, held-out forward test
window (see Section 4.5). Given its yearly resolution dataset (11 observations from 2015–2025 for the
evaluated locality), an 80/20 random train-test split was used (8 training and 3 test observations) for the real
estate model; because of the very small sample size at yearly resolution, this split is not chronological, and the
resulting test-set performance should be interpreted with this limitation in mind (see Section 4.3).
3.3 Supervised learning algorithms
Several supervised learning algorithms have been implemented to evaluate and predict investment returns
across multiple categories of assets. These include:
3.3.1 Structured data models
• Random Forest Regressor utilized for gold price prediction when nonlinear associations exist between
variables in a structured dataset. It is also used to forecast real estate prices by utilizing historical year-by-
year price trends for individual localities. For gold, the model used the past 10 years (2015–2025) of gold
price data as input features, with 100 decision trees (n_estimators = 100). The dataset was split
chronologically into records up to 2024 for training and beyond 2024 for testing to preserve temporal order
and prevent data leakage.
• Long short-term memory (LSTM) networks are applied to stock, mutual funds, and cryptocurrency data to
learn sequential patterns in time series data. Daily closing price data were fetched via the Alpha Vantage,
mftool, and CoinGecko APIs. Each model consisted of a single LSTM layer with 50 units and a dense output
layer trained using the Adam optimizer for 200 epochs. The look-back window was tuned to each asset's
data characteristics: a 90-day window for stock price prediction and a 60-day window for both mutual fund
NAV and cryptocurrency price prediction.
All the models were tuned for specific domains to improve the overall forecasting accuracy and flexibility of the
models to support the unified investment advisory framework.
3.4 Sentiment analysis module
In addition to quantitative forecasting, a sentiment analysis module was built to understand how news and
media coverage can influence investor perception across different asset classes.
For sentiment scoring, a lexicon-based approach is used where each token is matched against a list of predefined
positive and negative financial keywords. In addition, the ProsusAI/FinBERT model, which is a BERT model
specifically trained on financial text, yields more contextually accurate sentiment classifications than simple
keyword matching does.
The net sentiment score is calculated as follows:
Sentiment Score =
Positive Words−Negative Words
Total Words
(1)
Scores are classified as positive, neutral or negative. The positive, neutral and negative classifications in Fig. 1
represent the prevailing investor sentiment for a specific asset class or sector. The combined sentiment index is
intended to provide additional context for short-term price movements, complementing the quantitative
forecasts; a comparative evaluation of forecasting accuracy with and without sentiment features was not
performed as part of this study.
Fig. 1: Sentiment analysis output with positive and negative word counts.
3.5 AI advisory integration
AI Advisory uses a decision support tool that transforms large amounts of information into personal investment
advice presented as natural language. It uses the NVIDIA Nemotron Nano 9B V2 model provided by the
OpenRouter API to provide a bridge from raw analytical output to actionable user insights. The workflow is as
follows:
Input capture: A user requests investment advice via a JSON-formatted string (i.e., “Should I invest in TATA
Motors?”).
Query formatting: The query is converted into a formal financial format with the following structure:
Introduction; Market Overview; Pros and Cons; and Bottom Line.
Model request: The system sends a secure request to the model, requesting that it generate a markdown-
formatted response on the basis of real-time financial information.
Response processing: The system cleanses the response and formats it as a JSON-formatted string and displays
it on the dashboard.
This integration is intended to transform raw sentiment indicators and predictions into understandable
advisory statements, with the goal of increasing the accessibility of the platform's outputs for users who are not
able to interpret raw statistical metrics directly. This capability provides interactive, rationalized investment
suggestions alongside traditional numerical forecasting tools; a formal evaluation of its effect on user decision-
making or perceived credibility was not conducted in this study (see Section 5).
3.6 Data visualization
The proposed platform will present its analysis in two ways: tables and graphs. Graphs can also help create an
intuitive understanding of how an asset has performed historically and is expected to perform in the future by
displaying its past and future values on a line chart.
3.7 Graphical representation
The line charts are based on the date and net asset value (or price movement). The system shown in Fig. 2 uses
an interactive plot to show the data against net asset value (nav) and has a dark theme to allow better visibility
and readability.
Fig. 2: NAV trend visualization (Canara Robeco Large Cap Fund).
3.8 System architecture
The system architecture, as shown in Fig. 3, is focused on providing an integrated approach to combining
multiple data sources and utilizing an artificial intelligence-driven analytical model designed to provide real-
time, individualized and data-driven investment recommendations for all asset classes.
Fig. 3: System architecture of the proposed ai-driven unified investment guidance platform.
The platform is built around a Flask-based backend that acts as the central controller connecting all the
components. On the AI side, it manages three modules-the price prediction models (LSTM and Random Forest),
the FinBERT sentiment analysis module, and the Nemotron GenAI advisor that generates natural language
investment advice. On the data side, it reads user and portfolio information from an SQL database, pulls
historical price data from CSV files, and fetches live market prices from external APIs. All processed outputs are
then rendered and delivered to the user through a browser-based HTML interface.
For reproducibility, each of the two prediction model families (LSTM and Random Forest) is trained
independently on the corresponding asset-class dataset described in Section 3.1, using the training/test split
and hyperparameters detailed in Section 3.3, and is loaded by the Flask backend to serve predictions on request
rather than being retrained for every user query. When a user requests a forecast for a given asset, the Flask
controller identifies the corresponding asset class, retrieves the relevant historical data from the CSV files or
live prices from the external market APIs, applies the same pre-processing steps used during training-
chronological splitting for random forest-and min–max normalization and windowing for LSTM-and returns the
resulting prediction to the front end. The FinBERT sentiment module and the Nemotron-based GenAI advisor
operate as separate analytical services within the same backend: FinBERT scores incoming financial news
headlines using the process described in Section 3.4, while the Nemotron model is queried through the
OpenRouter API using the structured prompt format (Introduction; Market Overview; Pros and Cons; Bottom
Line) described in Section 3.5. User and portfolio information, used to personalize recommendations, is stored
in and retrieved from an SQL database, while all analytical outputs-price predictions, sentiment scores, and
advisory text-are combined and rendered through a Jinja-templated, browser-based HTML interface. This
modular design, in which each AI component is trained and served independently, allows the forecasting,
sentiment analysis, and advisory layers to be reproduced, retrained, or replaced without requiring changes to
the rest of the platform.
4. Results
4.1 Evaluation metrics
To assess the predictive performance of each forecasting model, three standard regression evaluation metrics
are used throughout this study: the Coefficient of Determination (R²), Root Mean Squared Error (RMSE), and
Mean Absolute Error (MAE). For a test set of n observations, where yᵢ is the actual value, ŷᵢ is the predicted value,
and ȳ is the mean of the actual values, these metrics are defined as follows.
The Root Mean Square Error (RMSE) measures the average magnitude of the prediction error in the same unit
as the target variable, with larger errors penalized more heavily:
RMSE = √((1/n) · Σ
i=1
n
(y
i
− ŷ
i
)
2
) (2)
The Mean Absolute Error (MAE) is a measure of the average absolute difference between the predicted and
actual values and is less sensitive to large outliers than the RMSE is:
MAE = (1/n) · Σ
i=1
n
|y
i
− ŷ
i
| (3)
The Coefficient of Determination (R²) is the proportion of variance in the actual values explained by the model,
where an R² closer to 1 indicates stronger predictive performance:
R
2
= 1 − [Σ
i=1
n
(y
i
− ŷ
i
)
2
/Σ
i=1
n
(y
i
− ȳ)
2
] (4)
In addition to these standard regression metrics, Section 4.7 also reports Directional Prediction-whether a
model correctly predicted the direction (upward or downward) of price movement over the evaluated test
window-to complement the error-based regression metrics above with a decision-relevant, direction-based
comparison.
4.2 Performance of stock forecasts by LSTMs
A predictive model is trained using historical stock price information from a representative subset of companies
listed on the National Stock Exchange (NSE) rather than a company-specific model so that a single generalized
model can be built across the many companies available. Using the Open, High, Low, Close, and Volume
information as input parameters, the model has a 90-day look back time frame to represent both short-term
and medium-term price movements. To avoid the possibility of overfitting, an early stop in the training process
is used. AAKASH. NS, the stock used for out-of-sample evaluation in this section, was not included in the training
set; the reported test performance therefore reflects the model's ability to generalize to a company it had not
previously seen, rather than performance on data used during training.
To assess the predictive performance of the model’s predictions over a one-month testing period (September
22nd, 2025 through October 22nd, 2025), the model’s predictions were assessed using common metrics for
regression models. For this model, the results were as follows:
• Root Mean Square Error (RMSE): 0.1487
• Mean Absolute Error (MAE): 0.1181
• Coefficient of Determination (R²): 0.78
In Fig. 4, we can see how well the model predicts the stock prices (orange dashed line) compared to the actual
prices of the stock (green line). Although some small variations in stock prices occur during sudden changes in
the market, it is clear that the model is able to track the direction of the stock price for the majority of the time.
This demonstrates the potential for the use of an LSTM architecture for the short-term prediction of stock prices.
The results show that LSTM models are capable of finding patterns within a dataset (such as temporal
dependencies or nonlinear relationships) that can be useful in predicting the near-term prices of stocks. As this
evaluation is based on a single out-of-sample stock (AAKASH. NS) over a one-month test window, these findings
should be regarded as a preliminary indication of the model's generalization ability rather than a general claim
of predictive performance across the broader stock market.
Fig. 4: AAKASH. NS stock future prediction trend.
4.3 Forecasting prices for real estate in Mumbai
Price forecasting for the Mumbai real estate market was completed using a random forest regressor trained on
yearly average price data for Andheri West from 2015 to 2025. The model used both the year and a nonlinear
feature (year²) derived from the year to reflect long-term trends of increasing prices. The optimal combination
of hyperparameters (number of trees, maximum tree depth, and minimum samples per split/leaf) was
identified using GridSearchCV with 3-fold cross-validation, optimizing for R².
Given the yearly resolution of the underlying dataset, the 11 available observations for Andheri West (2015–
2025) were split using an 80/20 random train-test partition (8 training and 3 test observations); the model was
fit on the training partition, and all reported metrics reflected the performance on the 3 held-out test
observations rather than the in-sample training fit. On this test set, the model achieved an MAE of 1330.74 INR,
an RMSE of 1434.79 INR, and an R² of 0.88, indicating that 88% of the variance in property prices on the held-
out observations is accounted for by the model. Given the very small test-set size, these results should be
interpreted as indicative rather than statistically robust.
The comparison of the predicted price trend from the fitted random forest model to the actual price trend for
Andheri West across the full 2015–2025 dataset period is shown in Fig. 5. The close alignment between the
predicted and actual trends supports the model's ability to capture the long-term real estate price trend, which
is consistent with the R² of 0.88 reported above. These results suggest that this method can be applied effectively
for modeling real estate price trends for specific urban submarkets.
4.4 Gold price forecasting
Random forest regression was used to predict the gold closing price on each trading day (GLD) using historical
data. To increase the accuracy of the forecasts, additional temporal information, such as the year, month, day
and price of the previous day, was also included in the model. Consistent with the chronological split described
in Section 3.3, the data were split into two parts: a training set up to 2024 and a testing set post-2024 to evaluate
how well the model would generalize.
The model's predictive performance was evaluated on the post-2024 test set using the regression metrics
defined in Section 4.1:
• Coefficient of Determination (R²): 0.92
• Mean Absolute Error (MAE): 0.9011
An R² of 0.92 indicates that the model explains more than 92% of the variance in daily gold closing prices,
reflecting strong agreement between the predicted and actual price series shown in Fig. 6. Building on this fit,
the model was also used to generate a 30-day forecast of future price trends (Fig. 7), allowing proactive decision-
making.
Fig. 5: Andheri West - Actual vs. Predicted price trend.
Fig. 6: Actual vs. predicted gold price (test set).
4.5 Cryptocurrency price prediction
To evaluate the ability of the proposed LSTM model to forecast cryptocurrency prices, historical Bitcoin price
data retrieved via the CoinGecko API (see Section 3.1) were used to train and test the model, following the same
LSTM architecture and chronological train/test split described in Section 3.3, with a 60-day look-back window
consistent with the window used for mutual fund NAV prediction. The model’s predictions were evaluated over
a test period from January 22, 2026, to March 22, 2026, and compared against a naïve forecast baseline to assess
the predictive value beyond a simple persistence model.
The model achieved a coefficient of determination (R²) of 0.80 on the test set, indicating that the LSTM captures
a substantial share of the variance in Bitcoin's price movements over the evaluation period, including the sharp
downward correction observed in early February 2026 and the subsequent partial recovery.
In Fig. 8, the predicted price (red line) is compared against the actual price (blue line) and the Naïve forecast
baseline (dashed gray line) for the test period. Compared with the naïve baseline, the LSTM model tracks the
overall downward and recovery trend of the actual price series more closely during the sharpest price
movements, particularly around the early February decline, although-which is consistent with the stock and
mutual fund results in Sections 4.2 and 4.6-some lag is visible during the most volatile short-term fluctuations.
This is consistent with the broader finding across asset classes that LSTM-based models are well suited to
capturing directional and medium-term trends in volatile, high-frequency financial time series, even where they
cannot fully anticipate sudden short-term shocks.
Fig. 7: Gold price future prediction trend analysis.
Fig. 8: Bitcoin price prediction using LSTM (test set).
4.6 Mutual funds prediction
To assess how well the proposed LSTM model performs in predicting mutual fund net asset value (NAV), we use
the Canara Robeco Large Cap Fund–Direct Plan (Growth) dataset. We used the historical NAV data to normalize
it to a range from zero to one using the min–max method. We also used a 60-day window in time to measure the
relationship between the input features and output features in terms of time dependence. We trained our model
using 80% of the data and tested it using the other 20%.
Our model performed satisfactorily in terms of prediction accuracy. Here are some of the key statistics about
how well our model did:
• R
2
score: 0.99
• MAE: 0.7398
• RMSE: 0.9575
In Fig. 9, we compare the actual NAV (blue line) to the predicted NAV (red line) for the test period. The red line
follows the blue line very closely in terms of the overall trend of the actual NAV; this finding demonstrates that
our model is able to learn and identify relationships between the input variables (the time series data) and the
output variable (the NAV).
Although minor prediction errors occur during periods of high volatility, the LSTM network effectively captures
the long-term growth trend of the mutual fund. Therefore, while we cannot claim to have developed the best
possible approach to mutual fund NAV prediction, these results demonstrate that deep learning models such as
LSTMs can be useful for financial time series forecasting, including mutual fund NAV prediction; this provides
insight into what may happen to the performance of a mutual fund moving forward.
4.7 Comparison with baseline investment strategies
To provide a preliminary assessment of the practical usefulness of the proposed LSTM model, its predictions for
AAKASH are presented. NS (Aakash Exploration Services Limited) were compared against two baseline
strategies over the test period (22-Sep-2025 to 22-Oct-2025): Naïve Forecasting and Simple Moving Average
(SMA-20).
Fig. 9: Mutual fund NAV prediction using LSTM (Canara Robeco large cap fund).
Table 7: Comparison of the proposed LSTM model against baseline investment strategies (AAKASH. NS)
Method
Directional prediction
Predicted end price (INR)
Actual end price (INR)
LSTM (Proposed)
Correct (Downward)
₹8.89
₹9.16
Naïve Forecast
Incorrect
₹9.93
₹9.16
SMA-20
Delayed
~₹9.50
₹9.16
As shown in Table 7, the proposed LSTM model correctly predicted the downward direction of AAKASH. NS over
the test period and produced a predicted end price of ₹8.89, compared with the actual end-of-period price of
₹9.16. This prediction was closer to the actual end price than those produced by the Naïve forecast and Simple
Moving Average (SMA-20) baselines were. The Naïve forecast, which carried forward the most recently
observed price, predicted an upward direction, whereas the SMA-20 baseline showed a delayed response to the
downward price movement. These results provide a preliminary comparison of the directional prediction
performance between the proposed LSTM model and the selected baseline strategies. However, because the
comparison is limited to a single stock and a one-month test period, the findings should not be generalized to
broader market performance. Further evaluation across multiple securities, asset classes, and longer test
periods is needed to assess the generalizability and practical applicability of the proposed approach (see Section
5).
5. Discussion
The experimental results across the five asset classes reveal a consistent pattern: model performance is closely
tied to the statistical structure of the underlying asset. The random forest and LSTM models applied to gold (R²
= 0.92) and mutual fund NAVs (R² = 0.99) achieved the strongest fit, which is consistent with the relatively
smooth, trend-dominated behavior of these series. The real estate model (R² = 0.88) performed similarly well
over most of the 2015–2025 period but showed larger deviations toward the end of the forecast window
(Section 4.3, Fig. 5), suggesting that longer-horizon forecasts of illiquid, low-frequency markets such as real
estate are more sensitive to unmodeled structural shifts than shorter-horizon forecasts are. In contrast, the
stock (R² = 0.78) and cryptocurrency (R² = 0.80) models-both applied to the most volatile and high-frequency
asset classes in this study-achieved comparatively lower predictive performance, with R² values indicating that
the models explain a smaller share of price variance than for the more structured asset classes. This finding is
consistent with the broader literature on financial time series forecasting, where higher-frequency, sentiment-
sensitive assets are inherently harder to predict than structured, slower-moving assets are.
[5-7,9,10]
This pattern has direct implications for the proposed platform’s design: rather than applying a single model
architecture uniformly across all asset classes, the results support the platform’s domain-tailored approach of
pairing random forest with structured, low-frequency data and LSTM networks with volatile, high-frequency
data (Section 3.3). Moreover, the comparison against baseline strategies in Section 4.7 shows that even when
absolute error metrics are moderate, as with stocks, the LSTM model correctly predicted the price direction
where the naïve and lagging moving-average baselines did not, indicating that the predictive value for real-
world investment decisions cannot be judged from R² or RMSE alone.
The integration of the FinBERT-based sentiment module and the LLM-based advisory layer (Sections 3.4–3.5)
addresses a limitation identified throughout the reviewed literature: that highly accurate models often sacrifice
interpretability, reducing investor trust in real-world use.
[1]
By converting raw sentiment scores and
quantitative forecasts into structured, natural-language advisory output, the platform aims to make model
outputs more actionable for retail investors without requiring them to interpret raw statistical metrics directly.
However, this interpretability layer has not been formally evaluated in this study-for example, through a user
study assessing whether the generated advisory text improves investors' decision-making relative to raw
forecasts alone-and represents an important direction for future validation.
Several limitations of the current study should also be acknowledged. First, the sentiment analysis module
currently draws on financial news headlines rather than a broader set of external indicators, such as
macroeconomic data, trading volume, or liquidity measures, that prior work has identified as omitted from
many existing models;
[2-8,13]
incorporating such indicators could further improve forecasting accuracy,
particularly for real estate and gold. Second, the baseline comparison in Section 4.7 is limited to a single stock
(AAKASH. NS) over a one-month test window; broader validation across a larger set of securities, asset classes,
and time horizons is needed before the platform's practical profitability can be more confidently established,
addressing a validation gap noted in prior literature.
[11,12]
Third, while the platform has been implemented and
deployed as a working system rather than remaining purely theoretical, addressing another gap identified in
the literature review,
[11]
this study has not yet evaluated the platform under live, real-time market conditions
over an extended period. Addressing these limitations-particularly through longer-horizon backtesting, broader
asset and instrument coverage, and formal evaluation of the advisory layer's usefulness to end users-represents
a promising direction for future work.
6. Conclusion
This study set out to address a specific gap in AI-driven financial technology: the absence of a unified platform
capable of guiding individual investors, particularly in India, across multiple asset classes rather than a single
market. To this end, in this research, an AI-driven Unified Investment Guidance Platform that integrates five
major asset classes-stocks, mutual funds, gold, real estate, and cryptocurrency-is proposed and implemented
within a single analytical and advisory framework, combining domain-tailored forecasting models (Random
Forest for structured, low-frequency data; LSTM networks for volatile, high-frequency data) with a FinBERT-
based sentiment analysis module and an LLM-based natural-language advisory layer. The experimental results
demonstrate promising predictive performance across the five evaluated asset classes, although performance
varies by asset type and evaluation setting: an R² of 0.92 for gold, 0.88 for real estate, 0.78 for stocks, 0.80 for
cryptocurrency, and 0.99 for mutual fund NAV prediction. Beyond these standard regression metrics, a
preliminary comparison against baseline investment strategies (Section 4.7) limited to a single stock over a one-
month test window revealed that the proposed LSTM model correctly predicted the direction of price movement
where the naïve and moving-average baselines did not. This is an encouraging early indicator of decision-
relevant value, although broader validation across more securities, asset classes, and longer horizons is needed
before general conclusions about the platform's practical investment performance can be drawn. The key
contributions of this work are the integration of five asset classes into a single platform, a forecasting
architecture tailored to the statistical characteristics of each asset class, a sentiment analysis module that draws
on diverse financial news sources, and an LLM-based advisory layer that translates quantitative and sentiment-
based outputs into interpretable, natural-language investment guidance. Taken together, these contributions
address several of the gaps identified in the literature review (Section 2.6), including the lack of unified
multiasset platforms, the trade-off between predictive performance and interpretability, and the shortage of
deployed, validated systems. From a practical standpoint, the proposed platform contributes to democratizing
financial guidance in India by providing individual investors-including those with limited financial literacy or
access to professional advisory services-a single, interpretable point of reference for comparing investment
opportunities across asset classes. This study has several limitations, discussed in more detail in Section 5,
including the platform's reliance on financial news sentiment rather than a broader set of macroeconomic
indicators, a baseline comparison limited to a single stock over a short test window, and the absence of a formal
evaluation of the advisory layer's usefulness to end users. Future work should extend the baseline validation to
a larger and more diverse set of securities and asset classes over longer time horizons, incorporate additional
macroeconomic and liquidity indicators into the forecasting models, and conduct a user-centered evaluation of
the AI advisory layer to assess its impact on investor decision-making in practice.
CRediT Author Contribution Statement
Mohammad Ali Ansari: Conceptualization, Investigation, Methodology, Software, Writing - Review & editing.
Huzaifa Zahid Husein Shah: Formal analysis, Investigation, Validation, Writing - Original draft. Talha Ansari:
Investigation, Visualization. Aditya Vijay Kamble: Data curation, Investigation. Ahlam Ansari: Supervision;
Writing - Review & editing. All authors have read and agreed to the published version of the manuscript.
Funding Declaration
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-
profit sectors.
Data Availability Statement
The datasets used in this study were derived from publicly accessible sources, including the Alpha Vantage API,
Google Finance (accessed via web scraping), CoinGecko API, AMFI (via the mftool Python library), MagicBricks,
Kaggle, The Economic Times, and SilverPrice.org, as described in Section 3.1.
Conflict of Interest
There is no conflict of interest.
Artificial Intelligence (AI) Use Disclosure
The authors declare that artificial intelligence (AI)-assisted tools were used only for language refinement,
grammar improvement, and manuscript structuring purposes during the preparation of this work. All technical
content, experimental implementation, results, and interpretations were independently developed and verified
by the authors.
Supporting information
Not Applicable
References
[1]
K. Mahajan, Transforming financial decision-making with artificial intelligence: a comprehensive study
on ai-driven algorithms for investment, trading, and portfolio management, Journal of Electrical Systems,
2024, 20, 5036–5044.
[2]
F. Maloku, B. Maloku, A. A. D. Kumar, A. Agarwal, House price prediction using machine learning and
artificial intelligence, Journal of Artificial Intelligence & Cloud Computing, 2024, 3, 1–10, doi:
10.47363/JAICC/2024(3)357.
[3]
N. Srikanth, M. Ramchander, Machine learning approach for house price prediction, Juni Khyat, 2023, 13,
81-88, available at: https://www.junikhyat.org.in/no_1_Online_23.html
[4]
K. Satyam, B. Suresh, Predicting real estate prices: A machine learning approach, International Research
Journal of Modernization in Engineering Technology and Science, 2025, 7, 5776–5780, available at:
https://www.irjmets.com/uploadedfiles/paper/issue_4_april_2025/73009/final/fin_irjmets17453126
00.pdf
[5]
H. M. Zangana, S. R. Obeyd, Deep learning-based gold price prediction: A novel approach using time series
analysis, Sistemasi: Jurnal Sistem Informasi, 2024, 13, 2581-2591.
[6]
A. Amini, R. Kalantari, Gold price prediction by a CNN-Bi-LSTM model along with automatic parameter
tuning, PLOS ONE, 2024, 19, e0298426, doi: 10.1371/journal.pone.0298426.
[7]
A. B. Nagata, M. S. Hidajat, D. A. Wibowo, W. Widyatmoko, N. B. Yaacob, Predicting gold price movement
using long short-term memory model, Journal of Applied Intelligent System, 2024, 9, 19-28, doi:
10.62411/jais.v9i1.10305.
[8]
X. Fan, J. Chen, Stock price forecasting in real estate industry based on investor sentiment, Frontiers in
Business, Economics and Management, 2022, 6, 54–59, doi: 10.54097/fbem.v6i3.3311.
[9]
S. Deepu, A. Verma, A. Raj, M. Kumar, S. Sen, K. Bharat, performance evaluation of real estate price
predictor using machine learning, Proceedings of the International Conference on Innovative Computing
& Communication (ICICC 2024), 2024, doi: 10.2139/ssrn.4856312.
[10]
S. Arslan, Bitcoin price prediction using sentiment analysis and empirical mode decomposition,
Computational Economics, 2025, 65, 2227–2248, doi: 10.1007/s10614-024-10588-3.
[11]
S. Tanwar, N. P. Patel, S. N. Patel, J. R. Patel, G. Sharma, I. E. Davidson, Deep learning-based cryptocurrency
price prediction scheme with interdependent relations, IEEE Access, 2021, 9, 143284–143295, doi:
10.1109/ACCESS.2021.3117848.
[12]
M.-F. Leung, A. Jawaid, S.-W. Ip, C.-H. Kwok, S. Yan, A portfolio recommendation system based on machine
learning and big data analytics, Data Science in Finance and Economics, 2023, 3, 152–165, doi:
10.3934/DSFE.2023009.
[13]
Deepthi C. G., J. J. Shetty, I. B. S. Jain, Lakshmi C. R. Kiran Kumar D. R., AI-powered finance management
platform, 2026 International Conference on Smart Futuristic Technology, Bengalore, India, 2026, 1-5,
doi: 10.1109/ICSFT66733.2026.11506666.
[14]
P. Jay, V. Kalariya, P. Parmar, S. Tanwar, N. Kumar, M. Alazab, Stochastic neural networks for cryptocurrency
price prediction, IEEE Access, 2020, 8, 82804–82818, doi: 10.1109/ACCESS.2020.2990659.
Publisher Note: The views, statements, and data in all publications solely belong to the authors and
contributors. GR Scholastic is not responsible for any injury resulting from the ideas, methods, or products
mentioned. GR Scholastic remains neutral regarding jurisdictional claims in published maps and institutional
affiliations.
Open Access
This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which
permits the non-commercial use, sharing, adaptation, distribution and reproduction in any medium or format,
as long as appropriate credit to the original author(s) and the source is given by providing a link to the Creative
Commons License and changes need to be indicated if there are any. The images or other third-party material
in this article are included in the article's Creative Commons License, unless indicated otherwise in a credit line
to the material. If material is not included in the article's Creative Commons License and your intended use is
not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this License, visit: https://creativecommons.org/licenses/by-
nc/4.0/
© The Author(s) 2026
