AI-based stock forecasts. How does it work? Technical indicators, AI & Machine-Learning.

Beyond closing prices, the model explicitly uses high and low information to capture intraday ranges, gap structure and extremes. Open, high, low and close values in multiple aggregations (e.g. rolling highs/lows over 5, 20 or 60 days) form the basis for trend, volatility and structure features.

Base quantities:
High(t), Low(t), Close(t)

Rolling extremes:
RollingHigh_n(t) = max(High[t−n+1 … t])
RollingLow_n(t)  = min(Low[t−n+1  … t])

The simple moving average over n days is used in multiple lengths: very short SMAs (5–10 days) capture short-term moves, medium SMAs (20–60 days) describe swing trends, and long SMAs (100–200+ days) serve as structural trend axes. Normalised distances between price and SMA (e.g. Close/SMA − 1) are also used as features.

SMA(t, n) = (1 / n) · Σ_{i=0}^{n−1} Close(t − i)

EMAs react faster to new information than SMAs. The model uses short EMAs (e.g. 8, 12 days) for short-term reversals, medium EMAs (26–50 days) for swing trends and longer EMAs (100+ days) for macro structure. Crossovers between EMAs of different lengths are modelled as separate structural features.

EMA(t) = EMA(t−1) + α · (Close(t) − EMA(t−1))
with α = 2 / (n + 1)

Beyond EMAs on prices, exponentially weighted means are applied to other features (e.g. volume, ROC, ATR). This produces smooth series that dampen outliers and highlight medium-term tendencies. Short EWM windows react more strongly, long ones produce more stable signals.

EWM_Feature(t) = EWM_Feature(t−1) + α · (Feature(t) − EWM_Feature(t−1))

MACD is used in several parameter variants: short pairs (e.g. 6/13/5) for very short-term momentum, classic settings (12/26/9) for mid-term signals and longer pairs (e.g. 19/39/9) for structural moves. The model also considers distances between MACD line, signal line and zero line.

MACD(t)   = EMA_fast(t) − EMA_slow(t)
Signal(t) = EMA(MACD(t), n_signal)

PPO normalises MACD relative to price level, making trends comparable across stocks with very different price ranges. The model uses short and medium PPO variants to capture percentage trend speed and strength.

PPO(t) = (EMA_fast(t) − EMA_slow(t)) / EMA_slow(t) · 100

Parabolic SAR produces a sequence of points often interpreted as dynamic stop levels. The model computes PSAR with different acceleration parameters to distinguish smooth trend phases from choppy markets.

PSAR(t) = PSAR(t−1) + AF · (EP − PSAR(t−1))
AF: Acceleration Factor, EP: Extreme Point

In addition to raw values, crossovers (e.g. golden/death cross between 50- and 200-day SMA/EMA) and time lags of indicators are used as separate features. This allows the model to learn how long a trend has persisted and how often reversals occurred within a window.

Examples:
ΔMA(t) = MA_short(t) − MA_long(t)
Lag(t, k) = Indicator(t) − Indicator(t−k)
CrossoverFlag = 1 if MA_short(t) > MA_long(t) AND MA_short(t−1) ≤ MA_long(t−1)

The Ichimoku cloud combines conversion line, base line and two leading spans projected into the future. The model uses the price position relative to the cloud, cloud thickness and the alignment of conversion and base line as structured trend features across shorter and longer windows.

Tenkan-sen, Kijun-sen, Senkou Span A/B according to the standard definition;
Cloud area = region between Span A and Span B.

RSI compares average gains and losses. In addition to the classic 14-day RSI, the model uses shorter (e.g. 7-day) and longer (21–28-day) variants to distinguish between short-lived spikes and persistent overbought/oversold regimes.

RS  = AvgGain / AvgLoss
RSI = 100 − 100 / (1 + RS)

The stochastic oscillator measures where the close lies within the recent high–low range. The model uses fast (%K) and smoothed (%D) variants across windows (e.g. 14, 20 days) to distinguish range-bound markets from trending regimes with frequent band breakouts.

%K = (Close − LowestLow) / (HighestHigh − LowestLow) · 100

ROC measures percentage price change between today and n periods ago. The model uses short ROC windows (e.g. 5 days) for short-term swings, medium (20 days) for swing moves and longer (60 days) to quantify larger trend moves.

ROC(t, n) = (Close(t) − Close(t−n)) / Close(t−n) · 100

CCI measures how far the current typical price (average of high, low, close) deviates from its moving average. The model uses shorter and longer CCI windows to distinguish brief spikes from structural extremes.

TypicalPrice = (High + Low + Close) / 3
CCI = (TypicalPrice − MA(TP)) / (0.015 · MeanDeviation)

Williams %R is similar to the stochastic oscillator but inverted (0 to −100). The model uses %R to spot fast reversal points in tight ranges and possible blow-off moves at trend extremes.

%R = (HighestHigh − Close) / (HighestHigh − LowestLow) · (−100)

The directional movement system provides ADX as a measure of trend strength and +DI/−DI for direction dominance. The model uses short ADX windows for current trend intensity and longer ones to gauge how long a trend regime has persisted.

Based on DM+ / DM− and True Range;
ADX ≈ smoothed measure derived from the difference between +DI and −DI.

ATR captures average trading range including gaps. The model uses short ATR windows (10 days) for current volatility, medium (20–30 days) for typical move sizes and long (60+ days) for structural risk regimes. ATR is considered relative to price and relative to its own history.

TR  = max(High − Low, |High − PrevClose|, |Low − PrevClose|)
ATR = (1 / n) · Σ TR

Bollinger bands adapt dynamically to volatility. The model uses several SMA/width combinations (e.g. 20 days / 2σ; 50 days / 2.5σ) to spot tight squeezes and sudden expansions. Time spent outside the bands and band width itself are additional features.

Middle = SMA_n
Upper  = SMA_n + k · σ
Lower  = SMA_n − k · σ

Keltner channels use an EMA as midline and ATR as band width. Unlike Bollinger bands they are less sensitive to single outliers. The model uses multiple EMA/ATR combinations to pick up breakouts from compressed channels and volatility expansions.

Middle = EMA_n
Upper  = EMA_n + k · ATR_n
Lower  = EMA_n − k · ATR_n

Donchian channels mark the highest high and lowest low of an n-day window. Short windows (e.g. 20 days) serve breakout systems, while longer windows (55+ days) represent higher-level trend boundaries. In the model, breakouts and pullbacks to these levels act as structural features.

Upper = max(High_{t−n+1 … t})
Lower = min(Low_{t−n+1  … t})

These volatility measures use high/low ranges and, for Garman–Klass, also open/close information. They provide finer risk signals than close-to-close volatility. The model uses short windows for current range turbulence and long windows for structural volatility regimes.

Vol_park and Vol_gar are based on High/Low/Open/Close using the standard Parkinson and Garman–Klass volatility formulas.

The model uses daily volume, moving volume averages and annualised metrics. Short windows (e.g. 5–10 days) capture current activity, longer windows (60+ days) reflect structural liquidity. Volume spikes relative to historical distributions are treated as separate features.

zScoreVol = (Volume − Mean(Volume_n)) / Std(Volume_n)

The money flow index combines typical price with volume and is often described as an 'RSI on money flows'. Short MFI windows show whether intraday moves are mostly bought or sold, while longer windows capture structural accumulation/distribution.

TypicalPrice = (High + Low + Close) / 3
MoneyFlow  = TypicalPrice · Volume
MFI        = RSI-like oscillator built from positive and negative money flows

Chaikin indicators combine the close's position within the daily range with volume. The model uses both the accumulation/distribution line and Chaikin money flow across windows to determine whether strong closes are genuinely backed by capital inflows.

AD = cumulative sum of
((Close − Low − (High − Close)) / (High − Low)) · Volume

On-balance volume cumulates volume depending on price up or down moves. The model uses normalised OBV variants (e.g. z-scores, percentage changes) to compare across stocks and detect divergences between OBV and price.

OBV(t) = OBV(t−1) + Volume(t) · sign(Close(t) − Close(t−1))

VWAP represents the volume-weighted average price. The model uses rolling VWAPs (e.g. 5-, 20-day), cumulative VWAPs since events (e.g. earnings) and anchored VWAPs starting at key swing points to approximate institutionally relevant price levels.

VWAP = Σ(Price · Volume) / Σ Volume

The model automatically detects higher highs (HH), higher lows (HL), lower highs (LH) and lower lows (LL) on multiple time frames. Transitions between bullish sequences (HH/HL) and bearish sequences (LH/LL) are labelled as break of structure (BOS). The current structural state (e.g. bullish, bearish, neutral) is encoded as a feature.

Rule-based detection of local highs/lows;
State transitions, e.g.:
• bullish when HH/HL sequence is active
• bearish when LH/LL
• BOS (break of structure) when the sequence type changes

Order block zones are identified via volume clusters, ranges and subsequent strong moves. Short order blocks capture intraday reaction levels, while wider, long-lived blocks likely reflect activity of larger market participants.

Cluster analysis of price–volume data to identify consolidation areas before strong moves (order blocks).

FVGs are gaps where price moved quickly with little overlapping trading in neighbouring candles. The model distinguishes between short-lived FVGs (e.g. news-driven) and long-standing imbalances that often get filled weeks later.

Bullish FVG: High(t−1) < Low(t+1)
Bearish FVG: Low(t−1)  > High(t+1)

The model combines structural volatility (e.g. ATR relative to zone size) with FVG and order block locations. This yields features that describe not only where a zone is but also how volatile the market typically behaves there.

Examples:
• ATR_in_zone / zone height
• Number of FVGs inside a zone
• Distance (in ATR units) from price to zone midpoint

Pivot points split the prior day’s range into potential support and resistance levels. The model uses classic, Fibonacci and Camarilla variants and considers daily, weekly and monthly pivots.

Classic:
P  = (High + Low + Close) / 3
R1 = 2P − Low
S1 = 2P − High
…
Fibonacci and Camarilla pivots use their respective multipliers.

Fibonacci retracements mark percentage pullbacks of prior moves (e.g. 23.6%, 38.2%, 50%, 61.8%). The model considers both retracements within local swings and on higher time frames to identify multi-layered reaction zones.

Retracement level = Start + (End − Start) · Fib factor
Fib factors: 0.236, 0.382, 0.5, 0.618, …