Automated technical chart analysis consistent, reproducible, scalable

Our forecasts are powered by an ensemble of modern machine-learning models specifically tuned for time-dependent financial data. The models process hundreds of technical indicators, market structure signals and volume information in parallel. Training is highly compute-intensive and runs on GPU-accelerated cloud infrastructure. Models are typically retrained or updated every night once new end-of-day data is available so the system can continuously adapt to current market conditions.

We work with high-quality historical market data, including closing prices, open, high and low prices, daily trading volume, and derived quantities such as returns, volatility and a wide range of technical indicators. Data is typically processed on an end-of-day basis after the market close, then cleaned and normalised. This ensures the models are built on consistent time series without distortions from outliers, splits or bad ticks.

The probability indicates how likely, according to the model, the price is to rise (UP) or fall (DOWN) over a defined horizon. For example, a value of 72% means that in historically similar situations the market went up in roughly 72% of cases. To avoid noise, we only show signals when probabilities are clearly elevated or clearly reduced, typically ≥ 70% for upward moves or ≤ 30% for downward moves. In the neutral zone between these thresholds the market is statistically hard to distinguish, so we deliberately avoid issuing signals there.

An UP signal means the model sees an elevated probability that the price will rise over the chosen horizon. A DOWN signal indicates an elevated probability of a price decrease. Both signals are purely statistical statements from technical analysis and are not direct buy or sell recommendations. They are intended as an additional building block within your own decision process, which should also consider risk profile, diversification and fundamentals.

The forecast horizon specifies over how many trading days the model’s expectation extends. A 1D horizon refers to the next trading day, 5D to roughly one week, 20D to several weeks and 60D to a mid- to longer-term period. Horizons are based on trading days, not calendar days. Very short horizons are much noisier, while longer horizons tend to reflect more structural trends and regimes.

Probabilities are not promises of a specific price outcome but statistical statements based on historically similar market situations. A probability of 70% does not mean the market will ‘certainly’ rise, but that it rose more often than it fell in comparable cases. Individual outcomes can always differ. Probabilities are best used as weighting or contextual input, not as guarantees.

Different horizons reflect different market dynamics. A positive 5D signal may indicate short-term momentum, while a negative 60D signal points to a broader downtrend. Such constellations are common during corrections or regime transitions. Many users use this information to adjust position size, stagger entries, or trade short-term setups more cautiously.

Signals are generated deliberately selectively. In calm or statistically ambiguous market phases, there may be extended periods without clear signals. During strong trends or high volatility, signals tend to appear more frequently. Exact frequency depends on the market, horizon and prevailing market dynamics.

Sideways markets and extreme crisis phases are among the hardest environments for any form of forecasting. In such phases, the separation between UP and DOWN scenarios often deteriorates, which is reflected in fewer or more cautious signals. The system is designed to surface this uncertainty rather than forcing artificial clarity.

Accuracy strongly depends on horizon, market regime and the specific stock. Financial markets always contain a large amount of randomness, especially in the short term. The goal of the models is therefore not perfect prediction, but a statistically measurable improvement over chance. What matters is whether probabilities are consistently better than neutral expectations across many cases — which is exactly what we measure with multiple metrics.

We strictly separate training, validation and test periods in time and rely consistently on out-of-sample evaluation. In addition, we use walk-forward testing, where models are repeatedly retrained and evaluated on subsequent, previously unseen data. This helps ensure that measured performance is not the result of hindsight optimisation but of genuine generalisation.

ACC (accuracy) describes in what fraction of cases the model predicted the direction correctly – i.e. how often UP and DOWN signals turned out to be right. BALACC (balanced accuracy) gives equal weight to up and down regimes so models do not look ‘good’ simply because they favour one trend direction. AUC measures the discriminative power of the model scores: cases that later move up should, on average, receive higher scores than those that move down. The Brier score evaluates the quality of probability forecasts: the lower the score, the better the predicted probabilities match realised outcomes. MCC (Matthews correlation coefficient) condenses all entries of the confusion matrix into a single robust number and is particularly useful for fair comparison of models across markets and time periods.

The hit rate measures how often a given signal turned out to be correct. For example, an UP signal with a high probability is counted as a ‘hit’ if the price rises over the chosen horizon, and as a ‘miss’ if it falls. Similarly, a DOWN signal must be confirmed by a subsequent price decline. The hit rate is the number of hits divided by the total number of signals. It can be computed separately for each horizon and provides an intuitive sense of how often signals have worked historically.

Walk-forward validation is a continuous evaluation approach for time series. Instead of training once on the full history and evaluating ‘in hindsight’, the model is advanced through time step by step: it is trained only on data that would have been available at that point, and then tested on the subsequent period. This produces a more realistic estimate of forecast quality across changing market regimes.

Out-of-sample means the evaluation is performed on periods the model had not ‘seen’ at training time. Our metrics are computed on real, continuously accruing market data as new trading days occur. This way, the numbers are not just an in-sample backtest, but an ongoing, timely measurement under real market conditions.

With financial time series, it is critical to keep features and labels strictly separated in time. We generate all input features only from information available up to the respective market close, and evaluate via walk-forward on subsequent periods. This prevents future information from unintentionally leaking into training, normalisation, or feature computation.

Forecasts are typically updated daily once new end-of-day data (after market close) is available. Signals are then ready before the next market open. This keeps analyses consistent and reproducible, relying on stable closing data rather than short-lived intraday noise.

Yes. For robust time series, price data and derived indicators are processed so common breaks (e.g., from splits) do not create artificial jumps in features and evaluations. The goal is a consistent history so signals and metrics remain comparable over long periods.

Stocks behave differently — liquidity, volatility, trend strength and reaction patterns can vary significantly. Also, a 1D forecast is statistically a different problem than a 60D forecast. That’s why models are trained per horizon and per asset, so forecasts better match the respective behaviour and time scale.

Financial markets are non-stationary and influenced by many external factors such as macroeconomic releases, interest rate decisions, geopolitical events and company news. Especially on very short horizons random noise dominates, so metrics like ACC or BALACC naturally tend to sit closer to the chance level. Lower values therefore do not automatically mean the model is ‘bad’, but often that the corresponding market segment was statistically hard to forecast. It is important to interpret metrics in the context of horizon, market regime and your own strategy.

Lags capture how past values of an indicator influence future price movements. For example, if a specific momentum signal three days ago has historically been followed by rising prices, the model can learn this lagged relationship and adjust the current UP probability accordingly. To make such relationships comparable across stocks with different price levels, volatilities and liquidity, all features are normalised. This prevents very expensive, highly volatile or illiquid names from dominating the model and ensures that probability scales remain consistent across the universe.

Instead of focusing on a single horizon, we combine multiple horizons into an overall measure. Short-term horizons are typically noisier and more volatile, while longer horizons capture structural predictive power more reliably. In many cases, longer horizons therefore receive slightly higher weight to produce a more robust overall picture, while shorter horizons still contribute information. The result is a metric that reflects both trading and investment perspectives and is well suited for comparing models or markets.

The forecasts are designed for informed retail investors and traders seeking data-driven decision support. They are particularly suitable for users who already work with technical analysis, fundamentals or systematic approaches and want to integrate probabilities into their own strategy.

No. Our forecasts do not constitute investment advice and do not replace personalised guidance from qualified professionals. They are automated, technically-oriented evaluations of historical price and indicator data. Your personal financial situation, risk profile and objectives are not taken into account. You should therefore treat the information as one building block within a broader decision and risk management framework.

For short-term setups, the 5D and 20D horizons are often more useful because they are less dominated by noise than pure 1D moves. In practice, signals work best as an additional filter — for example to confirm your own setup based on trend, support/resistance, volume, or a news context. It also helps to consider fundamentals and events such as earnings, and to manage risk via position sizing, stop levels, and diversification. Signals are statistical probabilities and should never be used as the sole decision basis.

For mid- to longer-term decisions, the 20D and 60D signals are often most relevant because they better reflect trend and regime phases. Many users apply these horizons to time entries and adds, to scale into positions, or to reduce risk during weaker phases. The most robust approach is combining signals with fundamentals (quality, valuation, growth) and a clear portfolio framework. Again, signals are not buy/sell recommendations — they are a data-driven complement to your own strategy.

We plan to offer new users a free one-month trial so you can explore the core forecasting features without any risk. After that, several options are conceivable, such as an inexpensive day pass for occasional access or a monthly / yearly subscription for regular users. Details of the final pricing model will be communicated transparently in the interface. In any case, the core goal is to provide high-quality, AI-powered technical analysis at a fair price.