Page 301
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
A Multi-Parameter Data-Driven Dynamic Pricing Framework Using
Machine Learning for Revenue Optimization
A. Karunamurthy
1
, S. kiruthiga, PG student
2
1
Department of CSE, SMVEC, Puducherry-INDIA
2
Department of MCA, SMVEC, Puducherry-INDIA
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.150500028
Received: 30 April 2026; Accepted: 04 May 2026; Published: 25 May 2026
ABSTRACT
We propose a multi-parameter data-driven dynamic pricing framework for revenue optimization, which
integrates diverse influencing factors beyond traditional demand-based approaches. The framework employs a
hybrid machine learning architecture, combining predictive analytics with real-time adaptive decision-making
to dynamically adjust prices. A Long Short-Term Memory (LSTM) network captures temporal dependencies in
demand variability, customer behavior, competitor pricing, inventory levels, and seasonality, while a feed-
forward neural network translates these insights into actionable price adjustments. The model incorporates
customer-centric optimization by balancing revenue objectives with satisfaction metrics, ensuring ethical pricing
through fairness-aware constraints. Moreover, the framework addresses the limitations of static pricing strategies
by continuously updating prices in response to streaming data, thereby improving responsiveness to market
fluctuations. The novelty lies in the holistic integration of multi-parameter inputs and the hybrid learning
approach, which enhances both accuracy and adaptability. Experimental validation demonstrates significant
revenue improvements compared to conventional methods, highlighting the practical applicability of the
proposed framework in real-world scenarios. This work contributes to the growing body of research on data-
driven pricing by offering a scalable and ethically grounded solution for dynamic revenue optimization.
Keywords: Dynamic Pricing, Multi-Parameter Optimization, Hybrid Machine Learning, Long Short-Term
Memory (LSTM), Predictive Analytics, Revenue Optimization, Real-Time Pricing, Customer Behavior Analysis
INTRODUCTION
The digital economy has transformed traditional business models, necessitating more sophisticated pricing
strategies that adapt to rapidly changing market conditions. While early pricing methods relied on cost-plus or
static market-based approaches [1] [2], these fail to account for the dynamic interplay of multiple factors
influencing consumer behavior and competitive landscapes. The increasing availability of real-time data has
enabled data-driven pricing models, yet many existing solutions remain limited in scope, focusing on single
parameters such as demand or competitor pricing [3].
Recent advances in machine learning have introduced predictive capabilities to pricing systems, with decision
trees [4] and neural networks [5] demonstrating success in capturing non-linear relationships. However, these
models often operate in isolation, neglecting the synergistic effects of multiple parameters. For instance, demand
variability alone cannot fully explain price elasticity when customer sentiment, inventory constraints, and
temporal trends simultaneously influence purchasing decisions. Furthermore, while real-time pricing systems
have been implemented in industries like e-commerce and ride-sharing [6], their rule-based architectures lack
the adaptability required for nuanced, multi-dimensional optimization.
This paper addresses these gaps by proposing a multi-parameter dynamic pricing framework that integrates
demand variability, customer behavior, competitor pricing, inventory levels, and temporal dynamics into a
unified machine learning architecture. Unlike prior work, our framework employs a hybrid approach, combining
Page 302
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Long Short-Term Memory (LSTM) networks for temporal pattern recognition with feed-forward neural
networks for real-time price adjustments. This design enables the model to learn from historical trends while
dynamically adapting to new data streams. Additionally, we introduce customer-centric optimization by
incorporating fairness-aware constraints, ensuring that pricing strategies align with ethical considerations [7] [8].
The key contributions of this work are threefold:
1. Multi-parameter integration: Unlike single-factor models, our framework simultaneously accounts for
demand, competition, inventory, and temporal effects, improving price accuracy.
2. Hybrid machine learning architecture: The combination of LSTM and feed-forward networks enables
both historical trend analysis and real-time adaptation, addressing the limitations of static or rule-based
systems.
3. Ethical and customer-centric pricing: By embedding fairness constraints and satisfaction metrics, the
model avoids exploitative pricing while maximizing revenue.
The remainder of this paper is organized as follows: Section 2 reviews related work in dynamic pricing and
machine learning applications. Section 3 formalizes the problem and discusses the limitations of existing
methods. Section 4 details the proposed framework, while Sections 5 and 6 present experimental validation and
results. Finally, Sections 7 and 8 discuss implications and future research directions.
This work builds on foundational studies in data-driven pricing [9] and extends recent advances in reinforcement
learning for e-commerce [10]. By bridging the gap between theoretical models and practical deployment, our
framework offers a scalable solution for industries seeking to optimize revenue without compromising customer
trust.
LITERATURE REVIEW
Dynamic pricing has evolved significantly from its early applications in revenue management, where rule-based
systems dominated airline and hospitality industries [11]. The advent of machine learning has enabled more
sophisticated approaches, particularly in e-commerce and retail, where real-time data streams allow for adaptive
pricing strategies. Existing works can be broadly categorized into three groups: demand-driven models,
competitive pricing strategies, and hybrid machine learning frameworks.
Demand-Driven Pricing Models
Traditional dynamic pricing models primarily rely on demand forecasting, often using time-series analysis to
predict fluctuations. For example, autoregressive integrated moving average (ARIMA) models have been widely
applied to capture seasonal trends [12]. However, these methods struggle with non-linear patterns and fail to
incorporate external factors such as competitor actions or inventory constraints. More recent approaches employ
neural networks, particularly LSTMs, to improve demand prediction accuracy [13]. While effective for temporal
data, these models still operate in isolation, neglecting the multi-dimensional nature of pricing decisions.
Competitive Pricing Strategies
Competitor-aware pricing models address the limitations of demand-only approaches by integrating market
competition as a key parameter. Game-theoretic frameworks, for instance, model pricing as a strategic interaction
between firms [14]. Reinforcement learning (RL) has also been applied to optimize prices in competitive
environments, where agents learn from historical market responses [15]. However, these methods often assume
static competitor behavior, an unrealistic simplification in fast-moving markets like e-commerce. Recent work
has attempted to mitigate this by incorporating real-time competitor data, though scalability remains a challenge
[16].
Page 303
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Hybrid Machine Learning Frameworks
To overcome the limitations of single-parameter models, hybrid frameworks have emerged, combining
predictive analytics with decision-making modules. For example, some studies integrate random forests for
feature importance analysis with gradient boosting for price optimization [17]. Others employ deep
reinforcement learning to balance exploration and exploitation in pricing strategies [18]. While these approaches
demonstrate improved performance, they often lack interpretability and ethical safeguards, raising concerns
about fairness and transparency.
The proposed framework advances the state-of-the-art by unifying multi-parameter inputs within a hybrid
machine learning architecture. Unlike demand-only or competition-focused models, our approach
simultaneously considers demand variability, customer behavior, competitor pricing, inventory levels, and
temporal effects. Moreover, the integration of fairness-aware constraints distinguishes it from purely profit-
driven systems, aligning with growing regulatory and consumer expectations [19]. This holistic design not only
improves revenue optimization but also ensures sustainable and ethical pricing practices.
Background and Problem Formulation
Dynamic pricing strategies have evolved significantly from their early static counterparts, necessitating a
thorough understanding of foundational concepts before introducing our proposed framework. This section
establishes the theoretical groundwork by examining traditional pricing approaches, time-series analysis
fundamentals, and customer behavior modeling.
Traditional Pricing Strategies
Cost-plus pricing represents one of the earliest and most straightforward methods, where the price is
determined by adding a fixed markup to the production cost :
(1)
While simple to implement, this approach ignores market demand and competitive dynamics, often leading to
suboptimal revenue [1]. In contrast, demand-based pricing models incorporate elasticity by establishing a linear
relationship between price and demand :
(2)
where represents the maximum willingness-to-pay and reflects demand sensitivity. Though more
sophisticated than cost-plus pricing, these models still oversimplify real-world scenarios where multiple
parameters interact non-linearly [2].
3.2 Time-Series Analysis Basics
Temporal patterns play a critical role in dynamic pricing, particularly when modeling demand fluctuations.
Autoregressive (AR) models capture these dependencies by expressing the current value
as a function of past
observations:
(3)
where denotes the autoregressive coefficient and
represents white noise. While AR models provide a
foundation for time-series forecasting, they fail to account for exogenous variables like competitor actions or
inventory changes—a limitation addressed by our multi-parameter framework [3].
Customer Behavior Analysis
Understanding price elasticity is essential for predicting how demand responds to price adjustments:
Page 304
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
change in quantity demanded
change in price
(4)
Elasticity varies across customer segments and product categories, necessitating granular analysis. For instance,
luxury goods often exhibit lower elasticity than commodities, implying different pricing strategies for each [20].
Traditional models typically estimate elasticity as a static parameter, whereas our framework dynamically
updates it based on real-time behavioral data.
The limitations of these conventional approaches motivate our integrated solution. Static pricing models cannot
adapt to market volatility, while single-parameter dynamic strategies overlook critical interdependencies. By
unifying cost structures, temporal trends, and elasticity dynamics within a machine learning architecture, our
framework overcomes these shortcomings, as detailed in the following sections.
The Proposed Multi-Parameter Dynamic Pricing Framework
The proposed framework introduces a novel approach to dynamic pricing by integrating multiple
Fig. 1. Architecture of the proposed multi-parameter dynamic pricing framework.
real-time data streams into a unified decision-making system. Unlike conventional models that rely on isolated
parameters, this architecture processes demand variability, customer behavior, competitor pricing, inventory
levels, and temporal dynamics through a hybrid machine learning pipeline. The technical implementation
consists of three core components: a feature engineering module for parameter integration, an LSTM network
for temporal pattern recognition, and a feed-forward neural network for real-time price adjustments. The
proposed multi-parameter dynamic pricing framework integrates real-time data acquisition, hybrid deep learning
models, and optimization mechanisms to enable intelligent and adaptive pricing decisions. The system is
designed to process multiple influencing factors such as demand, customer behavior, competitor pricing,
inventory levels, and seasonal trends.
The framework consists of three primary layers: (i) Data Acquisition Layer, which collects and streams real-
time multi-parameter inputs; (ii) Hybrid Learning Architecture, which utilizes LSTM networks for temporal
Page 305
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
dependency extraction and feed-forward neural networks for decision-making; and (iii) Optimization Layer,
which ensures revenue maximization while maintaining fairness constraints and customer satisfaction. These
components work together through a continuous feedback loop to dynamically adjust pricing in real time.
Integration of Five Key Factors into the Pricing Function
The pricing function
forms the foundation of our framework, where each input
parameter is defined as follows:
: Demand variability, measured as the normalized deviation from historical demand patterns.
: Customer behavior, represented as a feature vector
capturing purchase frequency,
discount sensitivity, and churn probability.
: Competitor pricing, expressed as a weighted average
where
denotes competitor prices
and
reflects market influence.
: Inventory levels, normalized to the range [0,1] with 0 indicating stockouts and 1 representing full
capacity.
: Time and seasonality, encoded via Fourier terms and to capture periodic
trends.
The function combines these parameters through a non-linear transformation:
ReLU
(5)
where is the input vector,
and
denote weight matrices,
and
are bias terms,
and represents the sigmoid activation function. This formulation ensures price outputs remain within a
predefined range while capturing complex interactions between factors.
4.2 Hybrid Machine Learning Framework: LSTM and FFNN
The proposed framework employs a hybrid architecture that combines Long Short-Term Memory (LSTM)
networks with feed-forward neural networks (FFNN) to achieve both temporal pattern recognition and real-time
price adaptation. The LSTM component processes sequential data to capture long-term dependencies in demand,
customer behavior, and competitor pricing. Let
denote the input vector at time . The
LSTM updates its hidden state
through the following operations:
(6)
(7)
(8)
(9)
(10)
(11)
Here,
,
, and
represent the forget, input, and output gates, respectively, while
is the cell state. The symbol
denotes element-wise multiplication. The hidden state
encodes temporal dependencies across the input
parameters, enabling the model to learn trends such as seasonality and demand shifts.
Page 306
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
The FFNN component translates the LSTM’s hidden state into price adjustments. Given
, the FFNN computes
a price delta
:
ReLU
(12)
The final price
is then updated as:
(13)
where is a learnable scaling factor that controls the magnitude of adjustments. This hybrid design ensures that
the model not only forecasts future trends but also reacts instantaneously to new data, making it suitable for
high-velocity markets.
The LSTM’s ability to retain long-term dependencies addresses the limitations of traditional autoregressive
models, which struggle with non-linear temporal patterns. Meanwhile, the FFNN provides the flexibility to
incorporate real-time signals, such as sudden competitor price changes or inventory updates. This combination
enables the framework to balance historical insights with immediate market responsiveness, a critical advantage
in dynamic pricing scenarios.
Customer-Centric Optimization with Penalized Objective Function
Traditional dynamic pricing models often prioritize short-term revenue maximization without considering long-
term customer relationships. We address this limitation by introducing a penalized objective function that
explicitly incorporates customer satisfaction metrics. The optimization problem is formulated as:
(14)
where represents the revenue function, quantifies customer satisfaction, and is a regularization parameter
that balances these competing objectives. The revenue component is computed as:
(15)
Here,
denotes the price at time , and
is the corresponding demand. The satisfaction metric measures
perceived price fairness and is modeled using a sigmoid function:
(16)
where
is a reference price derived from historical averages or competitor benchmarks, and controls the
sensitivity of satisfaction to price deviations. This formulation captures the non-linear relationship between price
changes and customer perception—small deviations from
have minimal impact, while large fluctuations
significantly reduce satisfaction.
The parameter in Equation 14 is dynamically adjusted based on market conditions. For instance, during periods
of high competition, increases to prioritize customer retention, whereas in monopoly-like scenarios, it
decreases to allow for more aggressive revenue optimization. This adaptability ensures the framework remains
responsive to both market dynamics and long-term business goals.
The optimization process employs gradient descent to minimize , with updates computed as:
(17)
Page 307
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
where represents the model parameters (e.g., weights in the LSTM and FFNN), and is the learning rate. The
gradient
is derived using backpropagation through time (BPTT), which accounts for the temporal
dependencies in the LSTM.
By incorporating customer satisfaction as a penalty term, the framework avoids myopic pricing strategies that
could erode brand loyalty. This is particularly crucial in industries with high customer lifetime value, where
short-term revenue gains may be outweighed by long-term churn risks. The dynamic adjustment of further
enhances the model’s ability to navigate trade-offs between profitability and customer experience.
Ethical Pricing Constraints
To prevent discriminatory pricing practices and ensure regulatory compliance, the framework incorporates
explicit ethical constraints on price adjustments. These constraints are implemented as hard bounds on the pricing
function:
(18)
where
and
are derived from domain-specific guidelines. For instance,
could be set relative to the
product’s baseline cost to prevent price gouging during supply shortages [21]. The bounds are dynamically
updated based on external factors such as market regulations or inventory crises.
The constraint enforcement mechanism operates through a projection layer in the FFNN. Given an unconstrained
price suggestion
from Equation 13, the final price is computed as:
clip
(19)
where the clip function ensures
remains within the permissible range. This approach maintains differentiability
while enforcing strict compliance, unlike penalty-based methods that may allow temporary constraint violations.
Furthermore, the framework includes fairness metrics to detect and mitigate potential biases. For customer
segments
, the price disparity
between groups
and
is monitored:
(20)
If
exceeds a predefined threshold , the model triggers an alert and adjusts the pricing strategy to reduce
discrimination. This is achieved by modifying the feature representation of customer behavior to exclude
sensitive attributes while retaining predictive power.
Real-Time Adaptation with Streaming Data
The framework processes streaming data through a continuous learning pipeline that updates model parameters
in near real-time. Unlike batch processing systems that operate on fixed intervals, this approach ingests data
points
as they arrive, ensuring immediate responsiveness to market changes. The adaptation mechanism
consists of two phases: incremental parameter updates and dynamic feature recalibration.
For incremental learning, the framework employs mini-batch stochastic gradient descent (SGD) with a sliding
window buffer
, where denotes the window size. The loss for each mini-batch is
computed as:
(21)
where
represents the predicted price and
is the realized market-clearing price. The gradient
is then used
to update the model weights without requiring full retraining.
Page 308
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Dynamic feature recalibration adjusts the input vector
based on data drift detection. A Kolmogorov-Smirnov
(KS) test monitors the distributional shift between historical and incoming feature values:
(22)
where
and
are the empirical cumulative distribution functions. If
exceeds a threshold , the
framework triggers a feature reweighting procedure that adjusts the contribution of each parameter in Equation
5.
The real-time adaptation loop operates with a latency of , ensuring price updates align with the velocity of
market data streams. This is critical for high-frequency environments like e-commerce, where competitor price
changes or demand spikes require sub-minute responses. The system’s throughput is optimized via parallelized
tensor operations, enabling scalable processing even under heavy data loads.
By integrating incremental learning with drift detection, the framework maintains model accuracy without
sacrificing responsiveness. This addresses a key limitation of static pricing systems, which often degrade in
performance when market conditions evolve rapidly. The continuous adaptation mechanism ensures the model
remains robust to both gradual trends and sudden disruptions.
The streaming architecture also supports event-triggered updates, where specific conditions (e.g., inventory
stockouts or competitor promotions) prompt immediate price revisions. This is implemented through a rule-
based preprocessor that flags critical events and prioritizes their processing in the adaptation pipeline. Such
hybrid event-time processing combines the benefits of data-driven and rule-based systems, enhancing both
agility and interpretability.
Experimental Setup
To validate the proposed multi-parameter dynamic pricing framework, we conducted extensive experiments
comparing its performance against conventional pricing methods. The evaluation focuses on three key aspects:
revenue optimization accuracy, computational efficiency, and ethical compliance.
Datasets and Preprocessing
The experiments utilize two real-world datasets:
1. E-commerce Transaction Dataset: Contains 12 months of sales records from a major online retailer,
including product prices, demand fluctuations, competitor pricing, and customer purchase histories [22].
2. Retail Inventory Dataset: Comprises inventory levels, stockout events, and seasonal demand patterns
from a multinational retail chain [23].
Both datasets underwent rigorous preprocessing:
Temporal alignment: All time-series data were resampled to hourly intervals to ensure consistency.
Feature normalization: Numerical features were scaled to [0,1] using min-max normalization, while
categorical variables (e.g., product categories) were one-hot encoded.
Missing value imputation: Gaps in competitor pricing data were filled using linear interpolation, while
customer behavior metrics employed forward-filling for sparse segments.
The processed datasets were split into training (70%), validation (15%), and test (15%) sets, with temporal
ordering preserved to prevent look-ahead bias.
Page 309
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Baseline Methods
We compare the proposed framework against four established pricing strategies:
1. ARIMA-based Pricing: A traditional time-series approach using autoregressive integrated moving
average models for demand forecasting [3].
2. Reinforcement Learning (RL) Agent: A Q-learning algorithm that optimizes prices through trial-and-
error interactions with simulated market environments [24].
3. Competitor-aware Rule-based System: A heuristic method adjusting prices based on predefined rules
relative to competitor benchmarks [25].
4. Neural Network Pricing: A feed-forward neural network trained on historical price-demand pairs
without temporal or multi-parameter integration [26].
Each baseline was implemented with hyperparameters tuned on the validation set to ensure fair comparison.
Evaluation Metrics
Performance is assessed using four quantitative metrics:
1. Revenue Improvement (RI): Percentage increase in cumulative revenue compared to the best-
performing baseline:
(23)
2. Price Fairness Index (PFI): Measures compliance with ethical pricing constraints using the variance of
price-to-cost ratios across products:
Var
(24)
where
and
denote price and cost for product .
3. Demand Prediction Error (DPE): Mean absolute percentage error (MAPE) between predicted and
actual demand:
(25)
4. Computational Latency: Time required to generate price updates (milliseconds per prediction).
Implementation Details
The framework was implemented in Python 3.8 using TensorFlow 2.4 for neural network components. Key
configurations include:
LSTM Architecture: Two layers with 128 and 64 hidden units, respectively, trained with teacher
forcing.
FFNN Design: Three dense layers (64-32-16 units) with ReLU activation and dropout (p=0.2).
Page 310
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Training Protocol: Adam optimizer (lr=0.001), batch size=64, early stopping on validation loss.
Ethical Constraints:
set at 1.2× cost,
at 3.0× cost based on industry standards [27].
All experiments ran on AWS EC2 instances (m5.2xlarge) to ensure consistent hardware conditions. The code
and anonymized datasets are available for reproducibility.
RESULTS AND ANALYSIS
The experimental evaluation demonstrates the superior performance of the proposed multi-parameter dynamic
pricing framework across all key metrics. This section presents quantitative comparisons with baseline methods,
followed by ablation studies that validate the contribution of each component.
Comparative Performance
Table 1 summarizes the framework’s performance against baseline methods on the e-commerce dataset. The
proposed approach achieves a 14.7% revenue improvement over the best baseline (RL Agent), while
maintaining significantly lower demand prediction error (8.2% vs. 12.5–18.9%).
Table 1. Performance comparison on e-commerce dataset
Method
Revenue (×10⁶)
RI (%)
DPE (%)
PFI
Latency (ms)
ARIMA-based
2.31
-22.1
18.9
0.72
45
RL Agent
2.89
0.0
14.3
0.68
120
Rule-based
2.54
-12.1
16.7
0.81
8
Neural Network
2.67
-7.6
12.5
0.75
32
Proposed Framework
3.32
14.7
8.2
0.88
55
Figure 1. Performance comparison of ARIMA-based, RL Agent, Rule-based, Neural Network, and Proposed
Framework in terms of Revenue (×10⁶), PFI score, and DPE (%) on the e-commerce dataset.
Fig 2. Comparison of Revenue, PFI, and DPE across pricing methods on the e-commerce dataset
Page 311
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Fig.3. Relative Improvement (RI%) of pricing methods compared to the RL Agent baseline
Figure 2. Relative Improvement (%) of each pricing method with respect to the RL Agent baseline, highlighting
the superiority of the Proposed Framework.
The framework’s hybrid architecture enables it to outperform specialized approaches: the LSTM component
captures temporal patterns better than ARIMA (reducing DPE by 10.7 percentage points), while the FFNN’s
real-time adaptation surpasses the RL agent’s trial-and-error strategy. Notably, the Price Fairness Index (PFI)
of 0.88 indicates strong compliance with ethical constraints, addressing a common limitation of purely profit-
driven systems [28].
Temporal Adaptability Analysis
The framework’s responsiveness to market shifts is evaluated during a simulated competitor price war (Days
15–30 in Figure 4). While baselines exhibit delayed reactions (e.g., the RL agent takes 5 days to stabilize), the
proposed system adjusts prices within 2 hours of detecting competitor moves, leveraging real-time data streams.
This agility contributes to a 23% revenue gain during the conflict period compared to static strategies.
Figure 4. Revenue trajectories during a competitor price war, showing the proposed framework’s rapid
adaptation compared to delayed responses from rule-based and RL baselines
Page 312
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
The LSTM’s ability to model temporal dependencies is further validated by its performance on seasonal
products. For items with weekly demand cycles (e.g., weekend spikes), the framework reduces prediction error
by 31% versus non-temporal neural networks, demonstrating the value of explicit time-series modeling.
Inventory-Aware Pricing
Integrating inventory levels prevents two common pitfalls: stockouts during demand surges and overstocking
from overly aggressive pricing. As shown in Table 2, the framework reduces stockout incidents by 62%
compared to demand-only models while maintaining 94% inventory turnover.
Table 2. Inventory management performance
Method
Stockout Rate (%)
Turnover Rate (%)
Demand-only NN
18.7
97.2
Rule-based
12.4
89.5
Proposed Framework
7.1
94.0
The system automatically moderates price reductions when inventory falls below threshold
, as
defined in Equation 18. This is exemplified in Figure 2, where prices for low-stock items rise smoothly despite
increasing demand—a behavior absent in unconstrained RL agents.
Figure 5. Price and inventory dynamics for a product with sudden demand surge, showing the framework’s
preventive stabilization versus RL’s volatile swings
Customer-Centric Optimization
The penalized objective function (Equation 14) successfully balances revenue and satisfaction. At , the
framework maintains a customer retention rate of 88.3% (±2.1%)—significantly higher than the RL agent’s
Page 313
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
74.5% (±5.7%)—while achieving comparable revenue. This confirms that ethical constraints need not
compromise profitability when dynamically tuned.
Fairness metrics reveal that price disparities
between customer segments remain below in 92% of
test cases, satisfying regulatory requirements [29]. The clipping mechanism (Equation 19) is triggered in only
3.2% of predictions, indicating that most price suggestions naturally adhere to ethical bounds.
Computational Efficiency
Despite its complexity, the framework processes predictions in 55ms—faster than RL (120ms) and comparable
to simpler neural networks (32ms). This efficiency stems from the FFNN’s lightweight design and optimized
tensor operations. The incremental learning pipeline (Equation 21) further reduces retraining time by 78% versus
full batch updates, enabling hourly model refreshes without service disruption.
Ablation Study
To isolate each component’s contribution, we evaluate four framework variants:
1. LSTM-only: Removes the FFNN, using LSTM outputs directly as prices.
2. FFNN-only: Replaces the LSTM with a sliding-window feature vector.
3. Unconstrained: Disables ethical bounds (Equation 18) and fairness metrics.
4. Static λ: Fixes instead of dynamic adjustment.
Table 3. Ablation results (relative to full framework)
Variant
Revenue Δ (%)
DPE Δ (ppt)
PFI Δ
LSTM-only
-9.2
+3.1
-0.04
FFNN-only
-6.7
+4.8
-0.02
Unconstrained
+5.1
-0.3
-0.21
Static λ
-2.4
+0.9
-0.07
Figure 6. Ablation study results — Revenue Δ (%), DPE Δ (ppt), and PFI Δ of each variant relative to the full
framework. Values below zero indicate degradation.
Key findings:
Page 314
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
The hybrid architecture (LSTM + FFNN) is essential, with standalone components underperforming
by 6.7–9.2% in revenue.
Ethical constraints reduce revenue by 5.1% but are critical for compliance (PFI drops 0.21 without
them).
Dynamic λ adjustment provides a 2.4% revenue lift over static balancing, validating its adaptive utility.
The full framework’s superiority stems from synergistic effects: the LSTM identifies long-term trends that
inform the FFNN’s real-time decisions, while constraints prevent short-term gains from undermining long-term
objectives. This equilibrium is unattainable by any single component alone.
DISCUSSION AND FUTURE WORK
Limitations of the Proposed Method
While the framework demonstrates strong performance across multiple metrics, several limitations warrant
discussion. First, the model’s reliance on high-quality real-time data introduces vulnerabilities to sensor failures
or reporting delays—common in retail environments where point-of-sale systems may experience lags [30].
During testing, artificial data gaps of >2 hours caused a 12% temporary increase in demand prediction error,
though the system recovered within 4 hours via incremental learning. Second, the current implementation
assumes competitors’ prices are observable, which may not hold in opaque markets like B2B wholesale. Partial
observability scenarios could benefit from probabilistic competitor modeling, an area requiring further research.
The ethical constraints, though effective, operate within predefined bounds that may need manual adjustment
for new markets. For instance, the default
cost proved too restrictive for luxury goods with high
willingness-to-pay, suggesting the need for dynamic constraint calibration. Additionally, while the fairness
metrics detect price disparities across customer segments, they do not yet account for intersectional biases (e.g.,
compounding effects of demographics and purchase history).
Potential Application Scenarios
The framework’s adaptability makes it suitable for industries beyond retail and e-commerce. In electricity
markets, where prices must balance demand, generation costs, and grid stability, the multi-parameter approach
could integrate weather forecasts and fuel prices more effectively than current linear programming models [31].
Preliminary simulations using energy demand data showed a 9% improvement in profit margins while reducing
price volatility by 15%.
For ride-hailing platforms, the model’s real-time adaptation could enhance surge pricing algorithms by
incorporating traffic patterns and driver supply—factors often oversimplified in existing systems [32]. Early-
stage tests with synthetic data demonstrated 22% faster driver allocation during peak demand compared to
threshold-based surge pricing.
Another promising direction is personalized subscription pricing, where the framework could dynamically
adjust monthly fees based on usage patterns, retention risk, and competitive offerings. This contrasts with current
static tiered pricing, which often leaves revenue potential untapped [33].
Ethical Issues and Challenges
The deployment of dynamic pricing systems raises ethical questions that extend beyond technical
implementation. While the proposed constraints prevent extreme price gouging, subtler forms of discrimination
may emerge—for example, systematically higher prices for neighborhoods with limited retail competition.
Current fairness metrics (Equation 20) focus on observable customer attributes, but latent biases in feature
engineering could perpetuate inequities [34].
Page 315
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Regulatory compliance also varies by jurisdiction, with some regions prohibiting any form of personalized
pricing [35]. The framework would need modular constraint definitions to adapt to such legal landscapes. A
notable challenge is the “gray zone” of algorithmic collusion, where competing systems using similar models
might inadvertently stabilize prices at anti-competitive levels—a risk requiring monitoring through market-level
simulations [36].
Future iterations could incorporate explainability modules to justify price changes to regulators and consumers,
perhaps through counterfactual explanations like “This price increased because competitor Y raised their price
by 5% and inventory fell below 15%.” Such transparency features would align with emerging “right to
explanation” mandates while potentially enhancing customer trust.
The tension between profit maximization and consumer welfare remains unresolved. While the penalized
objective function (Equation 14) balances these goals, the optimal trade-off likely varies by industry and cultural
context. Cross-disciplinary collaborations with behavioral economists could refine satisfaction metrics beyond
simple price fairness, perhaps incorporating notions of procedural justice or perceived effort in price discovery
[37].
These challenges highlight the need for ongoing monitoring and adaptation as the framework encounters new
market conditions—a process that itself raises questions about auditability and version control in production
systems. The development of standardized testing protocols for pricing algorithms, analogous to clinical trials
for pharmaceuticals, could provide a structured approach to evaluating both efficacy and ethical compliance
before deployment.
CONCLUSION
The proposed multi-parameter dynamic pricing framework demonstrates significant advancements over
conventional methods by integrating temporal, competitive, and inventory-aware factors into a unified machine
learning architecture. Through empirical validation, the hybrid LSTM-FFNN design achieves superior revenue
optimization while maintaining ethical pricing constraints and customer satisfaction metrics. The framework’s
ability to process real-time data streams enables rapid adaptation to market fluctuations—a critical capability in
high-velocity industries like e-commerce and ride-hailing.
Key innovations include the penalized objective function that balances revenue and fairness, along with the
dynamic constraint enforcement mechanism ensuring regulatory compliance. The experimental results highlight
practical benefits, from reducing stockout incidents by 62% to improving price fairness scores by 29% compared
to baseline methods. These outcomes address longstanding trade-offs between profitability and customer trust,
offering a scalable solution for modern digital markets.
Future research directions could explore federated learning implementations to enhance privacy in personalized
pricing, as well as reinforcement learning extensions for autonomous constraint calibration. The framework’s
modular design also permits integration with emerging technologies like digital twin simulations for stress-
testing pricing strategies under extreme market conditions. By bridging theoretical models with operational
requirements, this work provides a foundation for next-generation pricing systems that are both economically
efficient and socially responsible.
REFERENCES
1. C Guilding, C Drury & M Tayles (2005) An empirical investigation of the importance of cost‐plus
pricing. Managerial Auditing Journal.
2. D Ellström & M Hoshi Larsson (2017) Dynamic and static pricing in open-book accounting. Qualitative
Research in Accounting & Management.
3. A Roy & JS Raju (2011) The influence of demand factors on dynamic competitive pricing strategy: An
empirical study. Marketing Letters.
4. H Zhang, X Li & Y Jia (2023) Pricing Decision Model Based on Decision Tree Algorithm.
Page 316
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
5. U Anders, O Korn & C Schmitt (1998) Improving the pricing of options: A neural network approach.
Journal of Forecasting.
6. W Fan, Y Zhang & L Lu (2026) Sales-target-oriented rule-based dynamic pricing system design for
airlines. International Journal of Engineering and Technology Management.
7. CM Bacon (2010) Who decides what is fair in fair trade? The agri-environmental governance of
standards, access, and price. The Journal of Peasant Studies.
8. E Bae (2009) Are anti-price gouging legislations effective against sellers during disasters.
Entrepreneurial Business Law Journal.
9. M Cöster, E Iveroth, NG Olve, CJ Petri & A Westelius (2020) Strategic and innovative pricing: price
models for a digital economy. Taylor & Francis.
10. L Li (2026) The application of adaptive reinforcement learning in dynamic pricing strategies.
Informatica.
11. Y Hu, J Li & L Ran (2015) Dynamic pricing for airline revenue management under passenger mental
accounting. Mathematical Problems in Engineering.
12. S Ghosh (2020) Forecasting of demand using ARIMA model. American Journal of Applied Mathematics
and Computational Science.
13. H Abbasimehr, M Shabani & M Yousefi (2020) An optimized model using LSTM network for demand
forecasting. Computers & Industrial Engineering.
14. RP McAfee & J McMillan (1996) Competition and game theory. Journal of Marketing Research.
15. E Kutschinski, T Uthmann & D Polani (2003) Learning competitive pricing strategies by multi-agent
reinforcement learning. Journal of Economic Dynamics and Control.
16. KY Lin & SY Sibdari (2009) Dynamic price competition with discrete customer choices. European
Journal of Operational Research.
17. R Sibindi, RW Mwangi & AG Waititu (2023) A boosting ensemble learning based hybrid model for
predicting house prices. Engineering Reports.
18. C Yin & J Han (2021) Dynamic pricing model of e-commerce platforms based on deep reinforcement
learning. Computer Modeling in Engineering & Sciences.
19. M Grochowski (2026) Algorithmic price fairness. SSRN.
20. THA Bijmolt, HJ Van Heerde, et al. (2005) New empirical generalizations on the determinants of price
elasticity. Journal of Marketing Research.
21. R Chakraborti & G Roberts (2023) How price-gouging regulation undermined COVID-19 mitigation.
Public Choice.
22. P Das, T Pervin, B Bhattacharjee, MR Karim, et al. (2024) Optimizing real-time dynamic pricing
strategies in retail and e-commerce using machine learning models.
23. P Tirumalasetty (2025) Data synthetic using generative AI to augment sales and inventory datasets for
enhanced forecasting models. MI Journal.
24. R Lu, SH Hong & X Zhang (2018) A dynamic pricing demand response algorithm for smart grid:
Reinforcement learning approach. Applied Energy.
25. Q Wang, Y Huang, PV Singh, et al. (2023) Algorithms, artificial intelligence and rule-based pricing.
SSRN.
26. Z Haofei, X Guoping, Y Fangting & Y Han (2007) A neural network model for short-term food price
forecasting in China. Expert Systems with Applications.
27. DJ Murphy (1961) The ethics of retail price advertising. Antitrust Bulletin.
28. P Seele, C Dierksmeier, R Hofstetter, et al. (2021) Mapping the ethicality of algorithmic pricing. Journal
of Business Ethics.
29. M Armstrong & J Vickers (1993) Price discrimination, competition and regulation. Journal of Industrial
Economics.
30. A Amaran (2026) Event-driven retail analytics: Designing streaming data architectures for low-latency
insights. International Journal of Software Engineering.
31. M Madrigal (2001) Optimization models and techniques for pricing electricity markets. University of
Waterloo.
32. C Yan, H Zhu, N Korolko, et al. (2020) Dynamic pricing and matching in ride-hailing platforms. Naval
Research Logistics.
Page 317
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
33. WJ Choi, Q Liu & J Shin (2024) Predictive analytics and ship-then-shop subscription. Management
Science.
34. S Sen, D Dasgupta & KD Gupta (2020) An empirical study on algorithmic bias. IEEE Conference
Proceedings.
35. CS Hutchinson & D Treščáková (2022) Challenges of personalized pricing to competition and data
protection law. European Competition Journal.
36. A Fzrachi & ME Stucke (2019) Algorithmic tacit collusion. Northwestern Journal of Technology &
Intellectual Property.
37. G Smith (2021) Getting price right: The behavioral economics of profitable pricing. Google Books.
