From Charts to Circuits: How TINs Rewire Technical Analysis for the AI Era

TL;DR for operators

Trading platforms have spent decades giving users fixed technical indicators and then, more recently, neural models that treat those indicators as just another column in a feature table. Longfei Lu’s paper on Technical Indicator Networks, or TINs, proposes a different wiring job: make the indicator itself into the neural architecture.¹

That distinction matters. A TIN is not “MACD plus AI seasoning.” It maps the internal operations of an indicator—weighted averages, subtraction, smoothing, clipping, ratios, pooling—into layer operators whose initial weights reproduce the original formula. Before learning, the network behaves like the classical indicator. After learning, it can adapt while remaining traceable to the original calculation. Finally, someone remembered that explainability is easier when you do not begin by shredding the explanation.

The paper’s proof-of-concept is a MACD-based TIN trained with reinforcement learning on daily adjusted prices for US30 constituents. It tests traditional MACD, a price-only TIN-MACD, and a price-plus-OBV TIN-MACD. The headline result is not “AI beats the market.” The US30 buy-and-hold benchmark still has the highest cumulative return in the reported aggregate table. The stronger claim is narrower and more useful: the TIN variants, especially the OBV-augmented one, improve risk-adjusted metrics versus canonical MACD.

For business readers, the practical pathway is a platform architecture pathway. TINs could become adaptive indicator modules for broker dashboards, quant research tools, strategy libraries, and AI-assisted trading systems where users need both learning and auditability. The uncertainty is equally concrete: the paper does not include transaction costs, slippage, market impact, live execution, broad asset-class validation, or a full competition against production-grade trading systems. Treat it as a promising circuit diagram, not a money printer. The latter species remains mythical, despite what the internet keeps selling.

The familiar indicator is not the input; it is the architecture

The common way to combine technical analysis with machine learning is simple: compute indicators first, then feed them into a model. RSI, MACD, moving averages, Bollinger-style bands, volume measures—each becomes a preprocessed feature. The neural model then decides what to do with the table.

The paper argues that this approach leaves the indicator’s actual logic outside the model. The smoothing, differencing, normalization, and signal-generation mechanics are frozen before learning begins. The model may learn from the output of MACD, but it does not internalize MACD’s computational structure.

TINs invert that arrangement. Instead of placing the indicator beside the neural model, they rebuild the indicator inside the neural model. A moving average becomes a weighted linear layer. A MACD line becomes a topology of parallel average layers and a subtraction operator. A signal line becomes another weighted average layer applied to the MACD sequence. Other indicators can be decomposed into operations such as pooling, clipping, division, and normalization.

This is the paper’s central mechanism:

The important word is “initialized.” The TIN is not randomly born and then asked to discover MACD from scratch, which would be a heroic waste of compute and dignity. It begins with weights derived from the canonical indicator definition, so its pre-training output can reproduce the original indicator’s behaviour. Learning then adjusts the parameters inside a familiar structure.

That is the paper’s answer to a long-standing trade-off in financial AI: classical indicators are interpretable but rigid; black-box models are adaptive but harder to trust. TINs try to preserve the indicator’s semantic skeleton while giving it trainable joints.

MACD is the demonstration because it is simple enough to inspect

The paper uses MACD as the case study, which is sensible. MACD is not obscure, and its computation is naturally modular. It compares a slow moving average with a fast moving average, then smooths the resulting difference into a signal line. Traders typically read the crossover between the MACD line and signal line as a buy or sell signal.

In the paper’s formulation, a moving average over a window can be represented as:

The MACD component is then expressed as a difference between moving-average branches:

The exact sign convention is less important than the architecture: two smoothing branches feed a difference operator, and the resulting sequence feeds another smoothing branch for the signal line. In neural-network terms, this becomes a transparent computation graph, not a mysterious recurrent beast lurking in a basement.

The paper’s MACD TIN uses this topology in three steps.

First, the slow and fast moving averages become parallel linear layers. Their initial weights are set according to the chosen moving-average definition. For a simple moving average, the weights are uniform. For an exponential moving average, the weights follow an exponential decay schedule.

Second, the two moving-average outputs pass through a subtraction operator to form the MACD line.

Third, the MACD line passes through another smoothing layer to form the signal line, after which the system can generate trading actions.

That is the price-only version. The multidimensional extension adds auxiliary inputs while preserving the operator structure. In the experiment, the concrete auxiliary input is On-Balance Volume, or OBV: cumulative volume adjusted by the direction of price movement. The point is not that OBV is magically superior. The point is that TINs can extend a formerly univariate indicator topology into a richer input space without making the model completely unreadable.

Training adapts the circuit, but does not erase the diagram

The proof-of-concept trains the MACD TIN using reinforcement learning. The policy network is the TIN-MACD itself. It observes a state, chooses a buy or sell action, receives trading returns in the simulated environment, and updates parameters using a REINFORCE-style policy-gradient rule.

The paper also normalizes discounted returns within each batch and uses experience replay, where recent trajectories are retained and resampled for updates. These are implementation choices designed to stabilize learning and improve sample efficiency. They are not the main conceptual contribution. The main contribution is still the topology-preserving conversion of a rule-based indicator into a trainable network.

Each network processes a 52-day rolling window through a hidden layer of 26 units, aligned with the MACD parameterization $(12, 26, 9)$. The trading actions are generated from network outputs and executed using adjusted closing prices. Risk-adjusted performance is then measured with Sharpe and Sortino ratios.

The paper’s experimental progression is clean in intent:

This matters because the paper is not trying to prove that a TIN beats every possible LSTM, transformer, ensemble, or institutional trading stack. It explicitly frames such comparisons as conceptually awkward because those models often treat indicators as external features, whereas TINs are designed to internalize the indicator’s computation. That defence is fair up to a point. For proving the mechanism, comparing against canonical MACD is the right first move. For proving commercial trading superiority, the bar would be much higher. Conveniently, the market does not hand out revenue for elegant ontology.

The backtest supports feasibility, not a trading victory parade

The dataset consists of daily adjusted closing prices for the 30 US30 constituents, retrieved through yfinance. The paper reports an in-sample training period from 2015-08-25 to 2023-08-21 and an out-of-sample evaluation period from 2023-08-22 to 2025-08-22. Each equity is modelled independently with identical hyperparameter settings.

The aggregate results are worth reading carefully because they say something subtler than the usual “our model wins” chant.

The risk-adjusted story is favourable to TINs. The OBV-augmented TIN reports the strongest Sharpe and Sortino ratios. The price-only TIN also improves sharply on these ratios relative to traditional MACD.

The cumulative-return story is less flattering and more interesting. The US30 index has the highest cumulative return at 0.3461. The OBV-augmented TIN beats traditional MACD on cumulative return, but the price-only TIN is essentially tied with traditional MACD and slightly lower in the reported table. So the practical interpretation is not “TINs earn more in every sense.” It is “TINs may produce smoother return per unit of risk, especially when the indicator topology can absorb volume information.”

That is a better claim anyway. In real trading operations, raw cumulative return without risk, turnover, costs, and drawdown context is a fine way to impress amateurs and terrify risk managers.

The appendix provides per-constituent Sharpe and Sortino ratios and paired t-tests. The table includes the index as a separate target, giving $n=31$ in the paired tests. The reported Sharpe improvement for the price-only TIN is positive but not statistically significant: mean difference $+0.1574$, $p = 0.2113$. The price-plus-OBV variant is stronger: mean Sharpe difference $+0.3575$, $p = 0.0202$.

For Sortino, the price-only TIN reports a mean improvement of $+1.0371$ with borderline significance at $p = 0.0573$. The OBV variant reports $+1.0910$ with $p = 0.0485$. That sounds impressive, but the paper notes that when no downside deviation is observed, the Sortino ratio is set to 10.0. Several rows hit that value. This makes the Sortino evidence useful but more fragile than the Sharpe evidence. It is a signal, not a courtroom confession.

The appendix shows generality, but mostly as architecture sketches

The appendices broaden the framework beyond MACD. They map other indicators into TIN-style operator compositions:

This is where readers should avoid overbuying the paper. The appendices are valuable because they show that the TIN idea is not uniquely tailored to MACD. Many technical indicators really are compositions of familiar operations: weighted sums, differences, ratios, extrema, normalizations. That makes them good candidates for neural operator translation.

But most of the appendix indicator mappings are architectural demonstrations, not empirical validations. The only tested strategy family in the paper is MACD-based. A platform team could use the appendix as a design catalogue. A trading desk should not treat it as evidence that every classic indicator becomes profitable once rewired into PyTorch. That would be optimism wearing a lab coat.

The business value is adaptive explainability, not “AI trading magic”

For trading platforms, the commercial appeal of TINs is not that they replace every model. It is that they offer a middle layer between static charting tools and opaque AI systems.

Today, many retail and professional tools sit in two camps. One camp provides interpretable indicators that users understand but must manually tune. The other provides adaptive models that may perform better but are harder to explain, govern, and debug. TINs suggest a third product category: adaptive indicators whose internal structure remains tied to known financial logic.

That could matter in several business contexts.

For broker platforms, TINs could support “smart indicators” that begin as standard MACD, RSI, or CCI but adapt to asset class, volatility regime, or user-selected objectives. The selling point would not be guaranteed profit; it would be continuity. Users can see what the indicator is doing because its topology still resembles the tool they know.

For quant research platforms, TINs could become reusable modules inside strategy libraries. Instead of choosing between handcrafted rules and fully black-box models, researchers could test trainable versions of known signals with controlled topology and interpretable initialization.

For AI-assisted trading agents, TINs could provide structured tools that an agent can inspect, adjust, and combine. An LLM-driven agent that recommends a trading setup needs components with audit trails. A TIN offers a more legible object than “some hidden representation from a transformer said so,” which is not usually a satisfying sentence in a risk committee.

For compliance and model governance teams, topology preservation is the interesting bit. A model whose initial state reproduces a known indicator and whose parameters evolve within a mapped structure is easier to document than a generic neural model trained from scratch. Easier is not the same as easy. But in governance, “less awful” is a legitimate product feature.

What the paper shows, what Cognaptus infers, and what remains open

The paper directly shows that classical indicator operations can be reformulated as neural layer operators while preserving the original computational semantics. It demonstrates this in detail for MACD and sketches analogous conversions for several other indicators. It also reports a simulated MACD-TIN experiment in which the OBV-augmented TIN improves Sharpe and Sortino ratios over traditional MACD across the tested US30 setup.

Cognaptus infers that the strongest product opportunity is not autonomous trading but adaptive decision-support infrastructure. TINs are well suited to charting platforms, research workbenches, and controlled strategy-development environments where users want learning without losing interpretability. They are also natural candidates for hybrid systems where symbolic financial logic and trainable neural modules need to coexist without creating an explanatory swamp.

What remains open is the production case. The experiment excludes transaction costs. It does not model slippage, spread, market impact, borrow constraints, execution delays, order types, liquidity limits, or broker-side operational constraints. It uses daily data and relatively low trading frequencies, with holding periods described as days to weeks. That makes it less vulnerable to high-frequency execution noise, but it does not remove the need for realistic cost modelling.

The evidence is also narrow in model scope. MACD is the validated case. OBV helps in the tested variant, but the paper does not prove that every additional modality improves performance. Nor does it prove cross-asset robustness across futures, FX, crypto, commodities, or non-US equities. The author presents broader extension pathways, including richer inputs, advanced neural architectures, LLM-guided agents, broker APIs, and realistic market simulators. Those are sensible next steps. They are not results yet.

The sharper reading: TINs modernize the indicator layer

The best way to read this paper is not as a trading-strategy paper with a neural wrapper. It is an architecture paper with a trading proof-of-concept.

That distinction changes the evaluation. If the question is “Does this backtest prove live alpha?”, the answer is no. The evidence is not broad or realistic enough. If the question is “Can classical technical indicators be converted into trainable, interpretable neural modules without discarding their original logic?”, the answer is much closer to yes.

The mechanism is simple enough to be useful. Take the indicator apart. Map each operation to a layer operator. Initialize the weights from the canonical formula. Preserve the topology. Then train the parameters under a defined objective. That design gives practitioners a bridge from chart-based heuristics to adaptive models without demanding a leap into full black-box faith.

This is why the article should not be reduced to the performance table. The performance table is an early feasibility check. The architectural move is the real story.

TINs may or may not become a standard component of trading platforms. That depends on validation, implementation quality, cost-aware testing, and whether users actually want adaptive indicators that admit, inconveniently, that markets change. But the concept is well aimed: it modernizes the part of financial AI that has too often been stuck between rigid formulas and inscrutable networks.

Technical analysis has always been a kind of circuit diagram for market belief. TINs make that circuit trainable. The promise is not that the circuit will always be right. Markets remain rude like that. The promise is that when it learns, you can still see where the wires go.

Cognaptus: Automate the Present, Incubate the Future.