A system for the topological reconstruction of classical technical analysis indicators as modular neural networks for adaptive algorithmic trading systems: the Technical Indicator Network (TIN)

Abstract

A system for processing technical trading indicators using modular, topologically defined neural networks, comprising: - an input layer for capturing and preprocessing price data as well as structured and unstructured additional data from various sources, particularly those relevant to financial markets, including but not limited to trading platform API streams, news feeds, order book data, quantified sentiment vectors or other structured market data; - at least one indicator network (IN) whose layered structure reflects the topological reconstruction of the mathematical indicator logic; - one or more processing units as neural operation layers according to Figure 1 for aggregation, difference formation, normalization, limiting mechanisms, minimum / maximum determination, division and converted vector operations corresponding to the classical operators of technical indicators; - an output layer for generating one or more technical output values ​​for trading, classifiable as numerically representable decisions.

Description

1. Technical field

[0001] The invention relates to a method and a system for processing technical financial indicators using neural networks, in particular an adaptive, interpretable neural architecture model for mapping classical technical analysis indicators, including but not limited to moving average (MA), moving average convergence divergence (MACD), relative strength index (RSI), commodity channel index (CCI), rate of change (ROC) and stochastic oscillator, provided that their mathematical structure can be transformed into layer-based neural network topologies by transferable operators, as defined within the framework of the Technical Indicator Network (TIN system) introduced in this invention.

[0002] The technical field is an interdisciplinary area and includes, in particular, the following classes: • Algorithmic trading (algorithmic stock market trading), • Deep learning and neural networks, • Signal processing in financial markets, • Explainable artificial intelligence in the financial sector.

[0003] Specifically, the invention relates to the conception, construction, and use of so-called TIN systems, which are designed as formalized, interpretable neural network architectures for representing and extending rule-based technical analysis indicators. These TIN systems consist of standardized layer modules whose architecture and weighting are based on the mathematical structure of the original indicators.

[0004] The invention is based on the observation that many widely used technical trading indicators exhibit mathematical structures equivalent to layers of simple neural networks. Building on this, a framework is created in which each classic indicator is represented by a neural subnetwork (indicator network, IN) characterized by its topology, weighting matrix, and, if applicable, reinforcement learning method.

[0005] The invention therefore belongs to the fields of machine learning, deep learning modeling, and signal-based decision architecture in finance. In particular, it enables the operational use of classical trading indicators within modern AI-supported systems through structurally sound, explainable, and initializable neural networks.

[0006] Furthermore, the invention relates to applications in the areas of quantitative analysis, automated portfolio management, signal-driven trading execution, and the development of reusable AI building blocks for financial applications (financial informatics).

[0007] TINs represent a structured methodology for transforming technical indicators into scalable, reconstructible networks that can be extended through domain-specific functions. This makes it possible to dynamically adapt and train established indicators using artificial intelligence methods and integrate them into automated trading systems. The present invention thus exhibits both methodological and technical depth in the area of ​​network design and high relevance for practical applications in institutional trading systems where transparency, stability, and interpretability are essential. 2. State of the art and object of the invention

[0008] The current state of the art reveals various applications of deep learning models in the financial market environment, particularly long short-term memory networks (LSTM), convolutional neural networks (CNNs), transformers, and reinforcement learning for forecasting price movements and optimizing portfolios. However, in these approaches, technical indicators are usually processed as a mere input vector without considering their internal computational logic or structurally interpretable representation. Current practice is thus limited to a black-box approach to using technical indicators, neglecting their rule-based origins.

[0009] In practice, however, the following problems exist: Despite their origins in the last century, many classic technical trading indicators remain in active use today, providing robust, interpretable signals for various market regimes. Their rule-based structure and empirically validated effectiveness constitute a valuable knowledge base that has not yet been systematically utilized or developed further in modern AI systems. The challenge lies in translating this established knowledge into contemporary, machine-interpretable models without losing its proven mathematical structure. 1. Lack of structural representation of classical trading indicators within neural networks; 2. Low interpretability of the decision basis of neural trading models; 3. High optimization effort required when parameterizing technical indicators; 4. Low reusability or modularity in existing architectural designs; 5. Limited integration of additional data sources such as sentiment analyses, volumes, or cross-asset relationships into classic indicators; 6. Lack of reproducibility due to non-deterministic training processes; 7. High energy consumption and computational effort of modern deep learning models. The object of the invention is therefore to provide a technical method and a system in the form of a TIN system that enables the structural and semantic transfer of existing, functioning trading indicators into reconstructed neural networks. The original logic of the indicators is mapped by a topological network structure, and the initial weights are determined based on the original parameters. In this way, the functional core of classical methods is preserved and transformed into an extensible, adaptive network architecture, which serves as a methodological and structural basis for systematically converting existing rule-based trading indicators into trainable, explainable neural networks. This system does not represent a purely mathematical redefinition, but rather a technical architectural concept that builds upon proven indicator logic and leverages its structural and operational properties through the means of modern AI.

[0010] The TIN system enables: • the conversion of classical indicator functions into topologically defined neural networks; • Interpretable decision-making processes based on reconstructible layer structure; • Adaptive weighting optimization using reinforcement learning, supervised learning or other learning-based methods, especially taking into account real market environments such as volatility states, capital deployment requirements, drawdown limits or risk budgeting; • Modular expandability for additional data modalities (e.g. text data, order book data, event data); • A standardized, reusable architecture for the derived modeling of technical indicators based on defined operators (e.g., weighted averaging, difference calculation, division, MAD, minimum / maximum determination), as in Fig. 1 systematically classified.

[0011] This task addresses the technical problem of insufficient transparency, adaptability, and modularity when using technical indicators in the context of artificial intelligence and algorithmic trading. Complete backward compatibility with the original indicator logic is maintained, as many classic indicators can be interpreted from a network engineering perspective as special cases of neural networks with fixed weights. Simultaneously, the initial parameterization of the topologically represented network layers based on the originally defined indicator values ​​solves the central initialization problem of many deep learning models. This improves convergence, significantly reduces training effort, and allows the performance of classic models to be at least reproduced—and often surpassed—while maintaining algorithmic continuity with the existing trading logic.

[0012] For comparison: A TIN generated by the TIN system to represent a MACD indicator typically comprises only a few thousand trainable parameters, while a standard transformer model requires several million parameters. This results in a drastically lower computational and data requirement, making the TIN system not only technically efficient but also resource-saving and sustainable – especially in real-time applications with high frequency demands.

[0013] Furthermore, the reduction in training complexity and time has a direct impact on energy efficiency and cost-effectiveness in the development of production-ready trading systems. The TIN system thus enables a significant efficiency gain compared to complex deep learning models and is particularly suitable for real-time environments in high-frequency trading, where minimal latency, low computational load, and fast responsiveness are essential.

[0014] The invention creates a formal bridge between classical technical analysis and modern, AI-based model architectures through a systematic, technically interpretable mapping method of the indicator logic into neural network structures. 3. Description of the invention

[0015] The invention relates to a system for processing technical financial indicators, which is based on a modular neural network structure. This so-called TIN system transforms classic technical trading indicators—for example, MA, MACD, RSI, CCI, ROC, or stochastic oscillator—into reconstructible network layers. Each subnetwork (indicator network, IN) represents a specific indicator in a topologically defined network architecture. This architecture remains backward compatible with the original mathematical definition and allows the models to be used immediately without initial training.

[0016] An indicator network (IN) has at least the following components: • Input layer: Accepts sequential data streams such as price time series, volumes, order book data, sentiment vectors (e.g., generated by NLP models, rule-based classifiers, or other systems and processes), fundamental data, macroeconomic variables, or other data-driven sources, although this list is not exhaustive. The layer supports both structured and unstructured data sources in vectorized or preprocessed form and allows the integration of diverse data modalities from various origins. • Processing layers: The inner layer structure follows the mathematical logic defined by the respective technical indicator. It includes standardized operators according to Fig. 1. This includes weighted means, differences, divisions with bias regularization, limiting mechanisms, minimum / maximum determination, and statistical deviation calculations. The network topology is derived directly from the indicator's original formula. • Intermediate layers: For normalization, context integration, noise reduction, pooling, or other supporting transformations. Any layers commonly used in deep learning and integrable into the TIN system can be employed, including, for example, bias-regulated division, layer normalization, context filters, rescaling layers, attention mechanisms, or user-defined normalization methods. The selection of intermediate layers is not exhaustive and can be flexibly adapted to the specific indicator logic. • Output layer: Generates scalable numerical decision variables. These can be binary or categorical classifications (e.g., buy / hold / sell) or continuous regression outputs (e.g., valuation levels, risk scores). The decision is based on threshold logic, softmax classification, regression functions, or other suitable decision-making methods, depending on the use case and target metric.

[0017] The initialization of the weights is optionally: (a) analytically, by maintaining the original indicator structure and directly adopting the parameters specified in the original definition (e.g., 12-period moving average for EMA) for the initial weighting, or (b) adaptively, through reinforcement learning, supervised learning, or other machine learning methods based on historical data, market feedback, or capital target metrics.

[0018] The TIN system solves several technical challenges simultaneously: - it creates interpretability in neural models for financial indicators, - it significantly reduces training effort through structured initialization, - it enables complete reproducibility through standardized topologies.

[0019] Compared to generic architectures like LSTM, CNN, or Transformer, TINs have significantly fewer parameters (e.g., <10,000 instead of millions), drastically reducing energy consumption and hardware requirements. This makes the system particularly suitable for energy-efficient ("Green AI") applications in the algorithmic real-time trading environment.

[0020] Additionally, the TIN system enables the parallel integration of multiple indicators (INs) within a modular multi-indicator system. The outputs of multiple INs can be embedded into higher-level decision-making or analysis modules. This embedding can be aggregated, adaptive, or meta-network-based; weighted fusion is merely a special case.

[0021] The system allows for the direct transformation of existing technical analysis methods into AI-supported decision structures without altering the mathematical logic. This ensures that proven indicators remain fully usable, while the neural representation makes them adaptively expandable, explainable, and combinable. The modular coding allows for the simultaneous evaluation, historical testing, and combination of various indicators in live trading.

[0022] Integration into real trading systems (execution units, risk analysis modules) makes the system a transparent, resource-efficient building block for next-generation intelligent trading systems. The invention thus closes a technical gap in the application-oriented use of deep learning in the financial sector.

[0023] The present invention relates to the construction and application of a TIN system for the topological reconstruction of classical technical trading indicators, including moving average (MA), MACD, RSI, CCI, ROC and stochastic oscillator, as well as all other indicators that are structurally related to the operators defined in the TIN system according to Fig. The TIN system solves the problem of the lack of interpretability and adaptability of conventional indicator models by translating their mathematical structure into reconstructible network architectures. Each subnetwork (indicator network) corresponds to a technical indicator whose logical computational structure is based on a combination of these operators—including, but not limited to, weighted averages, differences, normalizations, divisions, limiting mechanisms, or minimum / maximum determination. The input layers support structured and unstructured data, particularly data relevant to financial markets. The weights can be either analytically initialized or dynamically adjusted—using a variety of learning methods, including, but not limited to, reinforcement learning.The system is designed for algorithmic stock trading, with a focus on transparency, explainability, and real-time adaptation. 5 examples

[0024] The following describes six exemplary TIN instances, each based on a classic technical analysis indicator. The input layer accepts any trading-related input data – this can be raw data (e.g., unprocessed price and volume flows) as well as preprocessed data (e.g., normalized price or volume values, technical indicators, or sentiment indicators) and is not limited to financial data, but can integrate any relevant time series data source to increase subsequent network stability and efficiency.

[0025] The implementation is achieved through topologically structured indicator networks (INs), which translate the mathematical structure and signal processing of the underlying indicators into modular neural architectures. The topological structures used are based on the combination of elementary operators, as described in Section 7. Fig. 1 are defined. This operator table maps the mapping between the original computational logic of the indicators and the functional processing layers of a neural network. 5.1 Example: TIN-MA (Moving Average) (see Fig. 2)

[0026] Network structure: - An input layer for receiving time-ordered price data, volumes, order book data, sentiment vectors, fundamental data, event information, or other relevant input data from various sources; - A moving aggregation layer that performs a weighted average over a defined time window. The weighting logic is based on fixed or dynamically adjustable parameters; - A comparison module that compares the aggregated average with the current or shifted market price and derives directional signals. This component can be implemented as a difference layer or a logical decision layer; - Optional: a time-delayed comparison structure for detecting average line crossovers. 5.2 Example: TIN-MACD (Moving-Average-Convergence-Divergence) (see Fig. 4)

[0027] Network structure: - An input layer as described above; - Two parallel-configured aggregation paths with weighted processing units to calculate a "fast" and a "slow" moving average; - A subtraction layer to determine the difference between the two moving averages, which generates the raw MACD signal; - A downstream smoothing layer that processes the difference signal; - An output layer to generate a classified trading signal based on threshold rules, difference direction, or crossover detection. 5.3 Example: TIN-RSI (Relative Strength Index) (see Fig. 8)

[0028] Network structure: - An input layer as described above; - Two separate aggregation modules for averaging positive and negative price changes, with filtering performed by ReLU-like separation functions; - A bias-regulated division unit for calculating the ratio of the means; - A normalization layer for transforming the result to a normalized range of 0 to 100; - An output layer for classifying the RSI level into threshold ranges. 5.4 Example: TIN-ROC (rate of change) (see Fig. 7)

[0029] Network structure: - An input layer as described above; - A subtraction module to determine the difference between the current and delayed price value over a defined time interval; - A division unit to normalize the difference with respect to the delayed price value, supplemented by regularization; - A downstream scaling unit to display the relative change value as a percentage; - An output layer for threshold classification. 5.5 Example: TIN-CCI (Commodity Channel Index) (see Fig. 10)

[0030] Network structure: - An input layer as described above; - A calculation layer to determine the typical price per period (arithmetic mean of high, low, and close); - An aggregation and averaging layer to calculate the moving average of the typical prices; - A downstream MAD layer to determine the average deviation from the moving average; - A division unit to normalize the deviation, supplemented by regularization; - An output layer to classify the CCI values ​​with regard to extreme zones. 5.6 Example: TIN stochastic (stochastic oscillator) (see Fig. 9)

[0031] Network structure: - An input layer as described above; - Two minimum / maximum modules to determine the local high and low points within a time window; - A difference calculation to determine the bandwidth and the distance to the minimum; - A division unit to normalize the current price level within the bandwidth (stochastic %K); - A smoothing layer to calculate the stochastic %D as a moving average of the %K value; - An output layer to interpret the %D signals based on predefined thresholds. 7 Description of the drawings Fig. 1: Operator table for processing layers Fig. Figure 1 shows a tabular overview of the operators used in the Technical Indicator Network (TIN) to map classic financial indicators, each with its corresponding mathematical formula: • 101 - Weighted Average Layer: A layer for weighted averaging over a time window of length N. Each processing unit (hereinafter referred to as a neuron) within the layer represents a specific weighted moving average with its own weights w. k,i , described by MAk,t=∑i=0N−1wk,i⋅Pt−i, ∑i=0N−iwk,i=1. In classic indicator design, the weight sums correspond to normalization to 1, but within the TIN system, they can be designed as learnable parameter structures. Examples of such indicators include: moving average (MA), moving average convergence / divergence (MACD), multiple moving average (MMA), commodity channel index (CCI), and relative strength index (RSI). • 102 - Subtraction / Addition Layer: A layer for calculating differences or sums of two input signals. Each neuron performs its own operation, described by Dk,t=Ak,t−Bk,t, Sk,t=Ak,t+Bk,t. Examples of indicators: MACD, MMA, CCI, Stochastic Oscillator. • 103 - Mean Absolute Deviation and Standard Deviation Layer: A layer for determining the mean absolute deviation (MAD) and the standard deviation (SD) over a window of length N. Each neuron calculates for its input signal: MADk,t=1N∑i=0N−1|Pk,t−i−P¯k|, SDk,t=1N∑i=0N−1(Pk,t−i−P¯k)2, with P¯k=1N ∑i=0N−1Pk,t−i, Example indicators: CCI, True Range (TR). • 104 - Multiplication / Division Layer: A layer for multiplication and division operations between input signals. Each neuron performs its own multiplication or division operation: Mk,t=Ak,t⋅Bk,t, Divk,t=Ak,tBk,t+ε. Examples of indicators: RSI, Rate of Change (ROC), CCI, Stochastic Oscillator. • 105 - Upper / Lower Bound Layer: A layer for limiting values ​​to defined upper and lower bounds O and U. Each neuron limits its input value individually according to TOPk,t=min(Xk,t,O), BOTTOMk,t=max(Xk,t,U). Example indicators: RSI. • 106 - Minimum / Maximum Layer: A summary layer for determining local extreme values ​​via a window. Each neuron determines the extreme values ​​for its input: MINk,t=mini=0,...,N−1Pk,t−i, MAXk,t=maxi=0,..,N−1Pk,t−i. Example indicators: Stochastic oscillator, support and resistance. Fig. 2: Topological structure of an MMA in the TIN Fig. Figure 2 shows the topology of a TIN for the MMA indicator within the TIN system: • 201 - Input layer: Captures sequential price data streams and forwards them to multiple parallel aggregation paths. • 202 - Weighted mean layer (fast): A weighted mean layer according to operator 101 ( Fig. 1), which calculates the fast moving average with a shorter time window. • 203 - Weighted mean layer (slow): A weighted mean layer according to operator 101 ( Fig. 1), which forms the slow moving average with a longer time window. • 204 - Subtraction layer: A difference layer according to operator 102 ( Fig. 1), which subtracts the result of the slow average from that of the fast average to generate the raw signal of the MMA. • 205 - Output layer (MMA signal): Consolidates the differentiated signal and forwards it as a trading signal to higher-level decision or aggregation modules. This description illustrates how the TIN system uses the classic calculation method of the MMA. MMAt=MAtschell−MAtslow=∑i=0Nf−1wi(f)⋅Pt−i−∑j=0Ns−1wj(s)⋅Pt−j transformed into a layer-based neural network structure. The mapping is done as follows: • The weighted sums MA schnell and MA langsam are implemented through the parallel layers 202 and 203 (operator 101). • The difference between the two partial values ​​is implemented by layer 204 (operator 102). This means that all constituent elements of the formula are implemented by modular, reconstructible layers, enabling an explainable and trainable implementation of the MMA indicator in the TIN system. Fig. 3: Example TIN with one-dimensional input data Fig. Figure 3 shows an example of a TIN that uses one-dimensional price data (301) or other trade-related time series data as input signals. This network structure is based on the one described in Fig. 2 MMA topology shown and uses the one in Fig. 1 specified operators for modular implementation. • 301 - Input layer for 1D data: Captures sequential price data points or other relevant input signals and forwards them to parallel processing branches. • 302 - Fast and slow weighted mean layer: Two weighted mean layers according to operator 101 ( Fig. 1), according to the in Fig. 2 defined MA structure. The initialization of the weights can be rule-based according to the original indicator or data-based using learning methods. • 303 - Subtraction layer: A difference layer according to operator 102 ( Fig. 1), which combines the results of the two mean layers to show the deviation between the price movements. • 304 - Multi-signal output layer: Consolidates the differentiated signal and, depending on the configuration, allows the generation of multiple numerical output values ​​(e.g., binary decision, regression value, risk assessment). This example structure illustrates how this can be done from Fig. The derived network can be extended for one-dimensional inputs. By combining the defined layers from Fig. 1 and the structured topology from Fig. 2. This network forms a scalable and trainable model for mapping technical indicators in a neural context. Fig. 4: Topological structure of a MACD indicator in the TIN system Fig. Figure 4 shows the network topology for representing the classic MACD indicator within a TIN. The structure is based on a sequence of moving averages and difference operations and utilizes the in Fig. 1 defined operators. • 401 - Input Layer: Captures sequential price data or similar trading information. • 402 & 403 - Weighted Average Layers (short / long): Two layers for calculating an exponential moving average (EMA). kurzand EMA lang , each according to operator 101 (weighted average). • 404 - Subtraction Layer: Calculates the difference between EMA kurz and EMA lang for forming the MACD curve; corresponds to operator 102. • 405 & 406 - Smoothing Layers: Additional weighted mean layers for smoothing the MACD curve (e.g. over 9 periods) and forming the signal line, according to operator 101. • 407 - Second subtraction layer: Calculates the difference between MACD curve and signal line to generate the MACD histogram, according to operator 102. • 408 - Output layer: Outputs the final MACD trading signal. Mathematical definition of the MACD indicator: MACDt=EMAtshort−EMAtlong Signalt=EMA9(MACDt) Histogram = MACDt signal

[0032] Illustration in the TIN system: • EMA kurz and EMA lang → Operator 101 (Layer 402 & 403) • MACD = Difference → Operator 102 (Layer 404) • Signal line = EMA via MACD → Operator 101 (Layer 405 & 406) • Histogram = Difference MACD - Signal → Operator 102 (Layer 407) This structure fully translates the mathematical logic of the MACD into a reconstructible neural architecture. The TIN system allows the weights of the involved EMA layers to be initialized with the coefficients defined in the classic MACD indicator. Furthermore, these parameters can be adjusted based on data without altering the interpretable structure of the indicator. Figure 5: 1D-TIN for the structured implementation of a technical indicator in the TIN system

[0033] Fig. Figure 5 shows a detailed implementation of a one-dimensional TIN as an example structure, in which the calculation of a classical indicator (e.g., MACD or MMA) is completely transferred to a neural network architecture. The in Fig. The 1 defined operators form the basis for the conversion of mathematical rules into structured layers. • 501 - Input layer (time-based price or volume data): Represents sequential input values ​​P t , P t-1 , ... over a window of N time steps. • 502 - MA langsam -Layer: Multiple parallel weighted average operations after operator 101 ( Fig. 1) with weightings wi,slangsam for slow smoothing. • 503 - MA schnell -Layer: Parallel weighted average operations after operator 101 ( Fig. 1) with weightings wi,kschnell for quick averages. • 504 - Subtraction Layer: Differentiating between MA schnell and MA langsam for each pair (according to operator 102), resulting in initial difference signals. • 505 - Aggregation / Fusion Layer: Consolidation of difference signals, e.g. using weighted sums or logical operations. • 506 - Final difference layer: Final output layer with aggregated subtraction (again operator 102) to generate the tradable signal as output.

[0034] This structure shows how the original mathematical components of a technical indicator are translated into layer operations: • MA structures → Operator 101 (weighted sums) • Difference formation → Operator 102 (Subtraction) • Structural fusion and smoothing → Combination of operators 101 and 102

[0035] The network is built entirely from standardized, reusable building blocks. The original calculation logic is retained, but is represented modularly by the TIN system, made capable of learning, and combinable with other indicator networks or data sources. Fig. 5 thus serves as an exemplary refinement of the in Fig. 3 topologies shown, specifically for one-dimensional applications. Figure 6: Expandable TIN (MACD-based) with 2D input data in the TIN system

[0036] Fig. Figure 6 shows a network based on the TIN system that processes two-dimensional input data (e.g., fundamental data, sentiment indicators, or text information). The underlying network topology is based on the MACD model from Fig. 4 and represents a scaled extension of the one-dimensional example Fig. 5. This structure illustrates the implementation of classic indicator logic not only on sequential price data, but also on higher-dimensional input data. • 601 - Input layer with 2D data: Matrix-like input data structure consisting of features such as financial ratios, sentiment indicators, company parameters or external event variables. • 602 - Weighted feature extraction: Several parallel weighted operators according to operator 101 ( Fig. 1) the vector structures aggregated for each data axis (analogous to MA) i,j ) calculate. • 603 - Subtraction Layer: Calculation of differences across multiple aggregation axes according to operator 102, e.g., between contrasting fundamental data or sentiment characteristics. • 604 - Fusion and smoothing: Combining the difference signals via adaptive weighted layers (operator 101) to structurally combine different dimensions of influence. • 605 - Output layer with decision vectors: Subtraction structure according to operator 102; optional classification logic outside of operators 101-106 that generates multidimensional outputs such as decision matrices, risk scores or dynamic thresholds.

[0037] This architecture illustrates how the TIN system extends an originally one-dimensional indicator structure to complex input data. Both the input and output layers can be designed to be multidimensional. The underlying indicator (e.g., a MA difference logic as in MACD) retains its mathematical structure and is transformed into a scalable, multi-channel architecture within the TIN system. Figure 7: Network structure of a ROC indicator in the TIN system

[0038] Fig. Figure 7 shows the network topology for mapping the classical ROC within the TIN system. The structure is based on weighted means, division and subtraction operators according to the [reference to be added]. Fig. 1 defined function blocks. • 701 - Input layer: Captures sequential input data over a time window N, for example, price data. The data source can also include other trading-relevant information such as volume, sentiment, or event vectors. • 702 & 703 - Aggregation Layers: Two weighted mean layers according to operator 101 ( Fig. 1) to calculate aggregated values ​​from different reference points, e.g. Pt and P t-N . • 704 - Division Layer: Operator 104 according to Fig. 1, to calculate the change quotient between the two aggregations. • 705 & 706 - Comparison Aggregations: Additional mean layer after operator 101 to calculate smoothed comparison values ​​(e.g. normalization, signal cleaning). • 707 - Subtraction layer: Difference operator according to operator 102 ( Fig. 1) to determine the final difference signal between the division result and the comparison value. • 708 - Output layer (ROC signal): Provides the final output value of the ROC indicator, for example for classifications or threshold checks.

[0039] Mathematical definition of the ROC indicator: ROCt=Pt−Pt−NPt−N=PtPt−N−1

[0040] Illustration in the TIN system: • Time comparison: Aggregations with operator 101 (Layer 702 & 703) • Rate of change: Division with operator 104 (Layer 704) • Signal smoothing or normalization: Operator 101 (Layer 705 & 706) • Differential signal: Operator 102 (Layer 707)

[0041] This network structure translates the mathematical definition of the ROC into a reconstructible layer topology. The original formula is preserved and technically implemented in a structured form within the TIN system. Figure 8: Network structure of an RSI indicator in the TIN system

[0042] Fig. Figure 8 shows the network structure for implementing RSI within the TIN system. The classical RSI logic is implemented using operators from Fig. 1. transformed into a structured architecture, where input and output data are not limited exclusively to price information. • 801 - Input layer: Captures sequential data streams (e.g., price, volume, text, or event vectors), not limited to the original RSI definition. • 802 - Aggregation Layer: Calculates aggregated data for preprocessing as a basis for limiting and filtering (Operator 101). • 803 & 804 - Upper bound / lower bound layer: Two separate bounding units for separating positive and negative components of the time series, according to operator 105. • 805 & 806 - Moving mean layer: Smoothing of the separated components using operator 101 (averaging), as input for strength analysis. • 807 - Division Layer: Ratio formation according to operator 104: RSIt = Average Gain / Average Gain + Average Loss • 808 - Output layer: Outputs the RSI value, e.g. as a continuous score, binary trading signal or decision vector.

[0043] Mathematical definition of RSI: RSIt=100−1001+RSt,withRSt=average gain average loss

[0044] Illustration in the TIN system: • Averaging gains / losses → Operator 101 (Layer 805 & 806) • Separation of positive / negative values ​​→ Operator 105 (Layer 803 & 804) • Forming ratios → Operator 104 (Layer 807)

[0045] This network structure transforms the RSI into a reconstructible subnetwork, while preserving the mathematical definition. The TIN system simultaneously allows for the integration of additional input data into the structure and the provision of outputs in various signal formats. Figure 9: Network structure of a stochastic oscillator in the TIN system

[0046] Fig. Figure 9 shows the topology of a subnetwork for representing the stochastic oscillator within the TIN system. The mathematical process is described using operators from Fig. 1. mapped and transferred into a structurable network format. • 901 - Input layer: Receives data sources such as prices, highs and lows, or external time series, e.g., P t , P high , P low The data may also include other trading-relevant characteristics such as fundamental data or external influencing factors. • 902 & 903 - Minimum / Maximum layer: Calculation of local extrema according to operator 106 to determine P low and P high to determine a time window. • 904 & 905 - Subtraction Layer: Operator 102 for calculating P high - P low . • 906 - Division Layer: Operator 104 for calculating the ratio between positional distance and total span: Pt−PlowPhigh−Plow • 907 - Smoothing layer: Mean layer according to operator 101 to generate the %D value by moving averaging the %K value. • 908 - Output layer: Calculates the stochastic oscillator signal, e.g. %K, %D or a derived structure.

[0047] Mathematical definition: %Kt=Pt−PlowPhigh−Plow,%Dt=MAn(%Kt)

[0048] Illustration in the TIN system: • Local extrema → Operator 106 (Layer 902 & 903) • Subtractions → Operator 102 (Layer 904 & 905) • Normalization → Operator 104 (Layer 906) • Smoothing → Operator 101 (Layer 907)

[0049] This network structure transforms the stochastic oscillator into a reconstructible layer topology. The original mathematical definition is retained and implemented technically in a structured form within the TIN system. Figure 10: Network structure of the CCI indicator in the TIN system

[0050] Fig. Figure 10 shows the network structure for mapping the Commodity Channel Index (CCI) within the TIN system. The computational steps of the classic CCI are thus translated into modular network layers. • 1001 - Input layer: Receives multidimensional input data (e.g., high, low, and closing prices, derivatives, or event data). The data sources can also include other trading-relevant variables. • 1002 & 1003 & 1004 - Mean Layer: Three moving mean layers according to operator 101 ( Fig. 1) They represent the typical price TP=H+L+C3 as well as its moving average MA TP . • 1005 - MAD Layer: Operator 103 for calculating the mean absolute deviation (MAD) of the TP from its MA. This module corresponds directly to the classical definition. • 1006 - Subtraction layer: Operator 102 for calculating TP t - MA TP the distance from the sliding center. • 1007 - Division Layer: Operator 104 for calculating the normalized CCI value: CCIt=TPt−MATP0.015⋅MADt • 1008 - Output layer: Outputs the CCI value or forwards it as input to downstream modules such as decision logic, classification or threshold evaluation.

[0051] Mathematical definition of classical CCI: CCIt=TPt−MATP0.015⋅MADt,TPt=Ht+Lt+Ct3

[0052] Illustration in the TIN system: • TP formation and smoothing → Operator 101 (Layer 1002 & 1003 & 1004) • Deviation calculation → Operator 102 (Layer 1006) • MAD calculation → Operator 103 (Layer 1005) • Normalization → Operator 104 (Layer 1007)

[0053] This structure allows for a fully reconstructible implementation of the CCI in the neural system, with extensibility for multiple data channels, dynamic weightings, and flexible decision logic. Layers 1004 to 1008 directly represent the original mathematical structure of the CCI, without altering the topology, and simultaneously provide the foundation for complex, context-adaptive applications. 8 Industrial Applicability

[0054] The invention is industrially applicable according to Section 5 of the German Patent Act (PatG) and Article 57 of the European Patent Convention (EPC), as it can be implemented in standardized software modules and used directly in operational trading systems. The claimed Technical Indicator Network (TIN system) has a deterministic structure that generates technical trading signals based on existing financial data streams. It enables the implementation of reconstructed classical technical indicators in the form of modular neural networks, whose weights can be either analytically initialized or adaptively trained.

[0055] The system is specifically designed for the following industrial applications: 1. Automated trading systems in the high-frequency domain (HFT): Due to its structural determinism and low computational load, the TIN system is suitable for use in real-time environments where latency sensitivity is a key requirement. The system's internal initialization with predefined weighting matrices significantly reduces training complexity, allowing the system to run on FPGA-, GPU-, or ASIC-based execution units. 2. Algorithmic strategy generation in institutional trading systems: The system can be part of commercial platforms for systematic strategy development, where automated decision proposals are generated based on technical analyses. Through the modular integration of multiple indicator networks (INs), the TIN system is suitable for multi-asset strategies in institutional portfolios. 3. AI-powered signal processing for broker and trading platforms: The system can be integrated into existing broker APIs, quantitative research platforms, or trading infrastructures, for example, as a neural backend service for the real-time generation of buy / sell / hold signals. The trainable parameter structures allow for dynamic adaptation to volatile market conditions without losing the interpretable structure. 4. Simulation and backtesting in quantitative analysis environments: The structurally accurate mapping of classical indicators in neural networks allows the system to be used in simulation and research systems for evaluating trading strategies. This also enables industrial applications in the areas of risk analysis, performance optimization, and strategy diversification. 5. Use in regulated trading systems: Due to the explainable architecture and the unambiguous topological mapping of the mathematical indicator logic, the system can meet regulatory requirements in the context of “Explainable Artificial Intelligence (XAI)” in the financial industry.

[0056] Implementation typically takes place in standardized programming environments (e.g., Python ecosystems with deep learning frameworks such as TensorFlow, PyTorch, or ONNX-compatible models) and is fully automatable. Compatibility with common financial data formats (e.g., Open-High-Low-Close-Volume (OHLCV) tables, order book data structures, or sentiment vectors generated by NLP systems) ensures seamless integration into existing industrial workflows. Therefore, the claimed system fulfills all requirements for industrial applicability within the meaning of Section 5 of the German Patent Act (PatG) and Article 57 of the European Patent Convention (EPC).

Claims

[1] A system for processing technical trading indicators using modular, topologically defined neural networks, comprising: - an input layer for capturing and preprocessing price data as well as structured and unstructured additional data from various sources, particularly those relevant to financial markets, including but not limited to trading platform API streams, news feeds, order book data, quantified sentiment vectors or other structured market data; - at least one indicator network (IN) whose layered structure reflects the topological reconstruction of the mathematical indicator logic; - one or more processing units as neural operation layers according to Fig.

1. Aggregation, difference formation, normalization, limiting mechanisms, minimum / maximum determination, division, and converted vector operations corresponding to the classical operators of technical indicators; - an output layer for generating one or more technical output values ​​for trading, classifiable as numerically representable decisions. [2] System according to claim 1, wherein the indicator network has a sliding aggregation structure based on time series operations and is suitable for moving average (MA) type methods. [3] System according to claim 1, wherein the indicator network comprises difference operations between aggregation branches with subsequent smoothing by weighted filter structures. [4] System according to claim 1, wherein division units with bias regularization are integrated to avoid numerical instability at small denominator values. [5] System according to claim 1, wherein the weighting parameters of the network layers are determined adaptively either by a structure-assigning mapping based on the original definition of the indicator or by training methods such as reinforcement learning, supervised learning or other machine learning methods. [6] System according to claim 1, wherein several indicator networks can be used in parallel and their outputs are embedded in a higher-level decision or analysis system, wherein the embedding can be aggregative, adaptive or sequential, including weighted fusion, voting mechanisms or metanetwork-based selection procedures. [7] System according to claim 1, wherein the input layer can process structured market data or unstructured text data from news feeds, trading platform API streams, trading volumes or quantified sentiment vectors in addition to price data, without being limited to these examples. [8] Training system within a TIN system for deriving technical trading strategies by targeted modulation of the weighting parameters in a neural network initialized by technical indicator logic, comprising: - an initialization unit that is applied to the reconstructed topological network layer of a technical indicator mapped within the TIN system; - a training module using reinforcement learning, supervised or other machine learning methods, which operates exclusively within the structural specifications of the TIN system and focuses on the technical adjustment of the weights within the indicator structure; - a simulated market interface consisting of feedback components for trading costs, position histories and state vectors linked to the TIN topology; - an evaluation system for deriving structured decision variables, which is directly linked to the indicator outputs of the respective TIN; - a weighting component, the parameters of which are determined either by a structure-assigning mapping based on the original definition of the respective technical indicators or by learning-based methods, for runtime-dependent relevance assessment; - a results fusion unit for signal combination and decision preparation.

Images