ML Model for Technical Indicators: RSI, MACD, and Bollinger Bands

A trading strategy based on RSI, MACD, or Bollinger Bands is classic, but fixed parameters fail. In a trending market, RSI can stay in overbought territory for weeks, while Bollinger Bands narrow, generating false breakouts. We build AI models that adapt these indicators to a specific asset and market state. The result is an increase in Sharpe ratio by an average of 1.5-2 times compared to standard settings, confirmed on out-of-sample data.

In real trading, simply using RSI with period 14 and thresholds 30/70 is insufficient. Even for a single instrument, optimal parameters change with market regime. We solve this through ML optimization: the model trains on historical data and dynamically selects indicator parameters. Additionally, we add contextual features: trend, volatility, volume. This avoids false signals during unstable periods. For example, in a volatile market, Bollinger Bands expand to 3σ, while in a sideways market they narrow to 1.5σ, improving breakout accuracy. Feature engineering is key: we generate over 50 features, including lags and cross-indicator products.

How ML improves RSI

RSI traditionally uses period 14 and thresholds 30/70. The problem: in a trending market, RSI can stay below 30 for weeks. The ML approach:

• Optimization of period (7-28) via Optuna for each instrument
• RSI in trend context: rsi_divergence_from_trend = rsi - trend_adjusted_rsi
• Multi-timeframe RSI: 7d, 14d, 21d as separate features
• RSI velocity: change in RSI over 3 days (acceleration signal)

We also apply neural network encoders to extract hidden patterns, but the baseline model is LightGBM, which ensures interpretability and speed.

Why combining MACD and Bollinger Bands yields better signals

MACD provides momentum signals, Bollinger provides volatility context. The ML model automatically learns when to use which signal.

Feature matrix — merging all indicators:

def build_technical_features(df):
    features = {}

    # RSI family
    for period in [7, 14, 21]:
        features[f'rsi_{period}'] = talib.RSI(df.close, period)
    features['rsi_divergence'] = compute_rsi_divergence(df)

    # MACD family
    macd, signal, hist = talib.MACD(df.close, 12, 26, 9)
    features['macd_hist'] = hist
    features['macd_hist_trend'] = hist.diff(3)
    features['macd_cross'] = (macd > signal).astype(int)

    # Bollinger
    upper, middle, lower = talib.BBANDS(df.close, 20, 2, 2)
    features['bb_pct_b'] = (df.close - lower) / (upper - lower)
    features['bb_bandwidth'] = (upper - lower) / middle
    features['bb_squeeze'] = (features['bb_bandwidth'] < features['bb_bandwidth'].rolling(126).min() * 1.1)

    # Stochastic
    slowk, slowd = talib.STOCH(df.high, df.low, df.close)
    features['stoch_k'] = slowk
    features['stoch_cross'] = (slowk > slowd).astype(int)

    return pd.DataFrame(features)

The model is LightGBM with quantile loss. Target: forward 5-day return. Feature importance reveals which indicators are truly predictive. We run Optuna for 100 iterations to tune hyperparameters (learning rate, max_depth, subsample).

Context dependency of indicators

Indicators work differently in different market regimes:

• Trending market: RSI < 30 = continuation of decline
• Ranging market: RSI < 30 = genuine reversal signal
• High volatility: Bollinger Bands need to be widened (3σ instead of 2σ)

Regime-conditional model:

regime = classify_market_regime(df)  # 'trending', 'ranging', 'volatile'

if regime == 'trending':
    signal = momentum_model.predict(features)
elif regime == 'ranging':
    signal = mean_reversion_model.predict(features)

Hull, J., "Options, Futures, and Other Derivatives"

Comparison of standard and ML approach

Results on real data

Process

1. Instrument: collect historical data, check quality.
2. Feature engineering: generate 50+ features based on RSI, MACD, Bollinger, Stochastic.
3. Optimization: use Bayesian search to tune indicator parameters.
4. Modeling: LightGBM with time-series cross-validation.
5. Backtesting: account for commissions, slippage, overnight swap.
6. Deployment: model in trading environment (SageMaker, Vertex AI, or your server).

Additional capabilities: model interpretability

Feature importance (SHAP) shows which indicator combinations actually work. For example, on S&P 500 futures test, top contributions are bb_squeeze (23%), macd_hist (18%), and rsi_divergence (15%). This allows us to filter out noisy features and simplify the model without quality loss. We also use partial dependence plots to validate model logic — it's crucial that dependencies are economically meaningful.

What's included in the work

• Baseline model with RSI, MACD, Bollinger, Stochastic
• Parametric optimization (Optuna/Bayesian)
• Regime detection (trending, ranging, volatile)
• Backtesting with real transaction costs
• Documentation and team training
• Code guarantee and 3-month support
• Target: Sharpe > 1.0 on out-of-sample data

Want a model adapted to your instrument? Get a consultation — we'll analyze the data and propose an architecture. Our engineers have 5+ years of experience in ML for financial markets and have completed 30+ projects building trading models. Contact us to discuss your task.

When does a time series forecasting model fail in production?

The CFO requests a quarterly sales forecast. An analyst builds SARIMA on three years of data, achieves MAPE 8.3% on the test set, and deploys. Two months later, the metric in production jumps to 23%. The root cause: the model was trained on pre‑COVID data, tested on a stable period, but production hit a promotion and supply chain disruption. Data leakage plus distribution shift—perfect notebook numbers, a broken forecast in reality. We have seen this pattern dozens of times across retail, fintech, and IoT. Our team has delivered more than 50 forecasting projects over 5+ years.

Incorrect cross-validation. Standard train_test_split for time series creates data leakage: the model sees future values during training. The correct approach is TimeSeriesSplit or walk‑forward validation with an expanding window.

Multiple seasonality. Hourly electricity consumption has three seasonalities: daily (24h), weekly (168h), yearly (8760h). SARIMA handles only one. Prophet can handle multiple but scales poorly to thousands of series.

Missing values and anomalies. A missing sensor reading is information (the sensor turned off), not NaN. Linear interpolation destroys this signal. Proper handling depends on the missingness mechanism.

Cold start. A new SKU in a 50,000‑item assortment has no history, yet a forecast is needed. Standard approaches fail; cross‑learning or feature‑based methods are required.

Why is model selection critical for your data?

Prophet (Meta) – a solid start for business data with clear seasonality and holidays. Fast setup, interpretable, built‑in outlier detection. Fails on irregular patterns and does not scale beyond ~10k series without parallelization.

Gradient boosting on features (LightGBM, XGBoost) – often underestimated. Engineer lags (t‑1, t‑7, t‑28), rolling means, day‑of‑week, holidays. The model trains on all series simultaneously, solving cold start via transfer learning. MAPE in retail often beats neural nets with proper feature engineering.

TFT (Temporal Fusion Transformer) – a transformer designed for interpretable forecasting with covariates. Built‑in variable selection, temporal attention, quantile outputs. Available in pytorch‑forecasting. Requires ~10,000+ records per series for stable training.

PatchTST – splits the series into patches (like ViT for images), capturing local patterns better than classic transformers. Excellent for long‑horizon forecasting (96–720 steps ahead).

N‑HiTS, N‑BEATS – attention‑free neural architectures, faster than TFT, competitive accuracy. N‑BEATS won the M4/M5 benchmarks for tasks without covariates.

How do we deploy TFT in production?

A typical pipeline via pytorch‑forecasting:

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="sales",
    group_ids=["store", "sku"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,  # 120 days
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,  # 28 days
    static_categoricals=["store_type", "category"],
    time_varying_known_reals=["price", "promo_flag"],
    time_varying_unknown_reals=["sales"],
    target_normalizer=GroupNormalizer(groups=["store", "sku"], transformation="softplus"),
)

A common mistake: the default target_normalizer (StandardScaler) breaks predictions for series with zero values (no sales on weekends). GroupNormalizer with transformation="softplus" is the correct choice for count data.

Case study: retail demand forecasting

A chain of 120 stores, 8,000 SKUs, 28‑day forecast horizon. The original system: SARIMA per series, MAPE 18.4%, retraining cycle – 6 hours. We replaced it with TFT on PyTorch + pytorch‑forecasting: a single model for all series, MAPE 11.2%, retraining – 40 minutes on an A10G. Feature importance via variable selection revealed that day_before_holiday influences more than the holiday date itself. Annual savings on inference alone exceeded $50,000.

Step‑by‑step configuration

1. Data collection and preparation. Handle missing values (mark NaN, interpolate only for technical failures), aggregate to required frequency, engineer covariates (holidays, promotions, prices).
2. Create TimeSeriesDataSet. Set group_ids (store + SKU), time index, forecast horizon. Choose target_normalizer based on target distribution.
3. Train a baseline. Prophet or LightGBM first – to understand complexity.
4. Train TFT. Use TemporalFusionTransformer with loss=QuantileLoss(), tune learning rate and hidden layer sizes.
5. Validate and interpret. Walk‑forward test, analyze variable selection, build attention heatmaps.

How to properly evaluate forecast quality?

RMSE alone is misleading – it over‑penalizes large values. Our standard set:

• MAPE – interpretable, unstable near zero.
• sMAPE – symmetric, avoids division by small numbers.
• MASE (Mean Absolute Scaled Error) – normalized relative to a naive seasonal forecast, ideal for comparing series of different scales.
• Pinball loss – for probabilistic forecasting, inventory management.

We guarantee a model card with these metrics on the validation set and walk‑forward results on at least 6 months of history.

What deliverables do you receive?

• Documentation of chosen architecture and hyperparameter rationale.
• Reproducible training and inference pipeline (Docker + CI/CD + Airflow/Prefect).
• Committed code with unit tests for key components.
• Team training: retraining, output interpretation, deployment of new versions.
• 3 months of post‑delivery support (consultations, bug fixes, fine‑tuning).

The model is deployed via FastAPI or Triton Inference Server. Retraining is scheduled (e.g., weekly) via Airflow with drift validation and automatic rollback if metrics deteriorate.

Process and timeline

We start with EDA: visualization, ADF test, STL decomposition, analysis of missing values and outliers. This takes 2–3 days but often reveals systemic data issues that block forecasting. Then we build a baseline (naive seasonal, Prophet), engineer features for LightGBM, and select a neural architecture if needed. Walk‑forward validation with a realistic horizon. Deployment via API with automatic retraining scheduled via Airflow or Prefect.

Timeline: MVP forecast on one data type – 3–6 weeks. Hierarchical forecasting system with automation – 2–5 months. Cost is calculated individually based on data volume, number of series, and required accuracy.

Our team consists of certified ML engineers (AWS ML Specialty, GCP Professional ML Engineer) with 5+ years on the market and over 50 completed forecasting projects. Contact us for a free analysis of your data – we will assess the task and provide initial recommendations within 1–2 days. Request a consultation to ensure your forecasts work in production, not just in a notebook.