Why an AI model for technical indicators breaks on regime change
RSI normalizes movement over a chosen window, and in a trend this works against the labeling: the indicator stays in the overbought zone for weeks, and the model reads each touch of the boundary as a reversal.
Some of these signals work out on a pullback, some do not, and it is impossible to distinguish them by a single value. MACD is even more honest in its weakness: it is calculated as the difference of moving averages, so signal lag is built into the formula itself.
The crossover is recorded when the move has already happened, and the histogram reverses later than the entry point.
Bollinger Bands add a third dimension — channel width. A band breakout with a compressed channel and a breakout with an expanded one are statistically different events, although for the model it is the same feature.
False indicator signals here are not noise, but a consequence of the fact that indicators describe price, not the cause of its movement.
The main break occurs during training. Overfitting to history is when an AI model for technical indicators finds combinations of RSI, MACD, and Bollinger Bands values that coincided with growth in a past window, but have no repeatable cause.
On historical data, accuracy looks high precisely because the same data participated in tuning, rather than testing it.
Analysis of RSI, MACD, and Bollinger Bands loses strength at the moment the character of the series changes: volatility, the ratio of trend and sideways, and relationships between instruments change.
Features selected on a calm segment stop predicting — and this is not weight degradation, but a shift in the data itself. We break down how this is addressed with regime-robust features and validation on data the model has not seen.
Signal accuracy, response lag, and robustness to regime change
Signal accuracy is not one number, but three independent dimensions: the share of false triggers, response lag, and behavior on regime change.
A model can win on the average historical segment and lose everything in a sideways market, where RSI reverses often and MACD gives false crossovers. Therefore, we accept work according to a fixed checklist, where each item is calculated separately on a holdout segment.
- Share of false triggers. We calculate not overall accuracy, but the share of entries closed by the protective condition before the target level — separately for each regime, not on average across the sample.
- Response lag. From the close of the bar with a confirmed signal to calculation of the target position, we keep fractions of a bar; otherwise the entry occurs at a price worse than the one at which the feature was calculated.
- Sideways market. A separate run on segments without directional movement: Bollinger Bands contract, and the model must not open an entry at every touch of the boundary.
- Trend segment. We check position holding through pullbacks — the signal must not reverse on the first return of price to the mean.
- Feature quality. We calculate RSI, MACD, and bands using a single methodology and a single smoothing period (TA-Lib as the reference) — a discrepancy produces a systematic signal shift.
- Regime-change robustness. We record the switching moment by volatility, not by a manual rule, and look at the error rate in the first hours after the change.
- Reproducibility. A repeated run on the same segment produces the same signals — without this, accuracy metrics mean nothing.
Design options: from threshold rules to hybrid schemes
Technical indicators provide a formalizable signal, so the way it is analyzed determines the robustness of the entire system.
Hard thresholds are human-readable and do not require accumulated data, but on a market regime change they produce a series of false triggers — RSI stays in the overbought zone longer than the rule assumes, and Bollinger Bands expand exactly when the previous threshold stops meaning anything.
| Option | When it fits | Limitations |
|---|---|---|
| Rules on indicators | Full explainability of the decision is needed, there is little history, the signal is checked manually | Breaks when volatility changes, thresholds have to be re-selected from scratch |
| Model on tabular features | Indicators are already reduced to a single snapshot at decision time, their nonlinear combinations matter | Loses event order: entering a zone and exiting it look the same |
| Sequence model | The indicator path itself matters — series, divergences, signal accumulation | Requires long history and careful labeling, support is more expensive |
| Hybrid scheme | Market regime changes, both explainability and flexibility are needed | Requires a regular retraining loop and distribution shift control |
The choice is made not by «model complexity» but by the amount of history, the cost of error, and the explainability requirement.
In practice, a hybrid scheme more often works: rules set the framework and cut off obviously unacceptable states, while the model refines the weight of features within the permitted area.
After that, its quality is maintained by a procedure: once per reporting period, regular retraining on a fresh data window, validation on a holdout segment, comparison with the current production model, and rollback if the metrics have dropped. Without this loop, any scheme degrades silently.
Implementation stages: from quote audit to operation
We proceed sequentially: first we check data suitability, then build features, and only after that train the model. Each stage is closed with a verifiable artifact — a report, a feature set, metrics. A skipped step in the data audit is most costly in operation.
-
Historical data audit. We bring quotes from different sources to a single time step, mark missing values and gaps in trading sessions. We check the series for look-ahead bias: the value of the current bar must not enter the calculation before the decision moment.
-
Feature design. We calculate RSI, MACD, and Bollinger Bands on rolling windows with lags, add derived features — price divergence from the indicator, band width, histogram sign change. For each feature, we fix the formula, window, and source.
-
Training and parameter selection. We split the series by time, not randomly; otherwise the model memorizes neighboring bars and shows inflated quality. We select parameters on a segment that does not overlap with the validation one.
-
Validation on a holdout period. We run the model on data it has not seen and separately examine behavior in different market regimes — trend, sideways, high volatility. If quality holds only in one regime, we redo the features.
-
Computation pipeline setup. We assemble regular recalculation of features and signals, compare the pipeline result with the training one on the same data: a discrepancy means an error in processing.
-
Handover to operation. We deliver the pipeline configurations, the retraining procedure, a dashboard with quality metrics, and data drift alerts. The client's team runs the calculation independently.
What you keep: code, features, procedures, documentation
The model is roughly half of the result; the other half is what allows it to be computed reproducibly and retrained without us.
Therefore, we hand over not only the AI model for technical indicators but the entire pipeline: code, features, calculation scheme, and operating rules, not just a single file with weights.
- Source code for training and calculation with pinned dependency versions — the environment comes up from scratch on a clean machine.
- Feature configuration: RSI windows, MACD periods, Bollinger Band deviations, rules for handling missing values and transitions between bar intervals — in declarative files, not hardcoded.
- Computation pipeline scheme: where indicators are calculated, where the model is applied, how data is updated, where checks and retries are placed on failures.
- Retraining procedure: trigger conditions — degradation of control metrics or a market regime change, training window, validation order, and rollback to the previous version.
- Control metrics on a holdout sample with fixed slices: versions are compared with each other on the same data.
- Technical documentation: description of features, input data formats, interfaces, and the model's applicability boundaries.
- Acceptance tests on reference data — with them, the calculation is reproduced on your side and compared with our result.
How we reduced the share of false signals in sideways markets: project breakdown
The client is an algorithmic trading team AlgoDesk: signals were generated automatically by RSI, MACD, and Bollinger Bands, but the pipeline did not distinguish trend from range.
In a trend, entries worked; in a sideways market, most trades closed by stop loss in the first minutes — price returned to the middle of the channel, although the indicators showed a consistent signal. Traders removed such entries manually, meaning automation did not reduce the workload.
We did not retrain the model on «successful» trades — that is overfitting to history. We added a separate regime assessment layer: based on volatility band compression, moving average slope, and divergence of confirming indicators, a segment is marked as trend or sideways.
In a range, the trigger threshold is higher and entry requires confirmation from a higher timeframe; in a trend, sensitivity remains the same so as not to miss the start of the move.
Metrics before and after
- Share of signals closed by stop loss immediately after entry in a sideways market: 61% → 23%.
- Average number of entries per hour inside a range: 14 → 5.
- Profitability on trend segments: within the previous range; entry delay — one candle.
- Drawdown in range-bound weeks: about three times lower.
- Manual signal cancellations by the trader: from dozens per shift to isolated cases.
Noise filtering did not kill sensitivity: stability in a trend was preserved, and false signals in a sideways market stopped being a stream. Controllability returned to the pipeline — the model marks an unfavorable regime itself, and the trader intervenes only in exceptions.
How does indicator analysis differ from forecasting on raw quotes?
Raw quotes are a noisy series: microstructure noise, gaps on news, volatility changes.
A model on such input learns to distinguish a pattern from a random spike, and it needs a long history and validation on a holdout period — otherwise it memorizes the past instead of the current regime.
Indicators are a compressed representation of the same series: RSI, MACD, and Bollinger Bands are already smoothed and normalized, so the input is more compact and robust to single outliers.
Indicator analysis wins where explainability of the result matters. A MACD crossover, RSI exiting the overbought zone, a touch of the channel boundary — the decision decomposes into verifiable conditions that can be run through history and discussed with risk control.
There are fewer false triggers on short windows, and the amount of data required for training is noticeably more modest.
Forecasting on raw quotes is needed when the signal lives in the structure of the series itself: a volatility regime change, a prolonged trend, a reaction to a news gap. Indicators lag here — they are built on smoothing and physically cannot keep up with a sharp change.
For such tasks, the model is trained on a price sequence, and indicators are connected as additional features, not as the only source.
We choose the approach based on the task, not habit: horizon, acceptable signal lag, amount of history, explainability requirements. In practice, a hybrid more often works — calculation by indicators as the framework and a model on the series itself for subtle dependencies.
Explainability is preserved: it is visible which indicators contributed and which parts of the series strengthened the signal.
Retraining, data leakage, and gaps: common concerns
Three objections come up at almost every start: how to prove that the model is not overfit to history, what to do with missing values in the series, and whether the result will fall apart on a market regime change. Short answers on the merits — without them, it is too early to discuss architecture.
How do you check that the model is not overfit to history?
We split the series by dates, not by random sampling: training on one segment, validation on the next, then the window shifts — the model passes through different market states. Features are calculated only on closed bars; otherwise the indicator peeks into the future.
We look not at the best result but at the spread between windows: if it stays in a narrow corridor, there is no overfitting.
What should be done with gaps and missing values in the series?
RSI, MACD, and Bollinger Bands are calculated over a continuous window, and a missing value inside breaks the value.
We align the calendar by trading intervals, not by calendar days; we restore short gaps, and cut long ones and mark them as a gap so that the model does not learn from a splice of heterogeneous segments.
How do you understand that the model does not work in only one market regime?
We calculate metrics separately for trend and sideways: average accuracy hides a failure in one of them. The report shows where the signal is confident and where the model gives a cautious result. A regime change is visible in the feature distribution in advance, not from account drawdown.
How explainable is the result?
For each signal, it is visible which indicators contributed and at which values — down to the historical segment that is recalculated manually. This is also the acceptance criterion: the client's engineer reproduces the signal and gets the same number.
Let's discuss the task and come back with an estimate and work plan
It is enough to bring what you already have: an export of historical quotes for the required instrument, the signal calculation rules in their current form, and a description of how the result is used now.
This is enough to estimate the data composition, stages, and scope of work — a separate specification at the input is not needed.
- Data composition: sources and history depth, frequency, known missing values and outliers, which indicators we calculate (RSI, MACD, Bollinger Bands) and with which averaging windows.
- Pipeline requirements: where the model is computed — a closed internal pipeline or external services, who has access to data and results, whether training and inference need to be separated.
- Target behavior: what we consider a correct signal, what forecast horizon, which is more expensive — a false trigger or a missed move.
- Operating mode: one-off calculation, regular scheduled recalculation, or continuous stream, and with what delay from candle appearance the result is needed.
- Degradation risk management: how the model is switched off, what we consider quality deterioration, and by which set of metrics it is visible.
- Acceptance points: on which historical segment we validate, who on the client side confirms the result, and how we document a discrepancy.
- Access and roles: who is responsible for exports, who for acceptance, whether test data with restricted access is needed.
After the discussion, we will come back with a timeframe estimate, a stage-by-stage work plan, and requirements for data and the pipeline — in writing, so they can be agreed within the team.
