Hummingbird Probabilities Mapping

October 7th, 2025

This is the fist version with updates to be released within the month. Many thanks for your support and I hope that this EA opens doors and becomes highly profitable for you. Cheers!!

What the Hummingbird Probabilities Mapping EA is....

A nature inspired EA that utilizes combinations of the following trend patterns and projects a probability mapping with greater than 70% accuracy based on real-time analysis.

EMA Trend
MACD
RSI
VWAP Spread
Burst
Squeeze
Volatility (ATRp)

-> Nature-inspired, probabilistic fusion of multi-indicator features into a single "state" output
-> States: hover (neutral), dart (long bias), backstep (short bias), perch (risk-off)

The Hummingbird EA blends several classic technical features into a single probabilistic “intent” read on the market—long, short, neutral, or risk-off—then trades only when a new bar closes. It’s designed to be:

• Non-repainting for trades (decisions occur on closed bars).

• Robust (attaches even when markets are closed or history is thin, and keeps showing a status HUD).

• Readable (probabilities are displayed as whole-number percentages).

The high-level loop (every tick or timer “tick”)

1. Collect recent bars with CopyRates(...) into rates[] . This avoids direct use of Close[]/Time[] and works reliably for EAs.

2. Detect the latest closed bar ( rates[1] ). The EA always evaluates the market from this bar so the HUD updates constantly, even if a new bar hasn’t formed yet.

3. Build features from standard indicators (EMA/RSI/ATR/Bollinger/MACD), microstructure signals (burst, squeeze), and a session VWAP with daily resets.

4. Fuse features → scores → probabilities via a temperature-scaled softmax: you get pLong , pNeut , pShort (0–100% on the HUD).

5. State machine sets one of four “hummingbird” states:

   • dart → long bias

   • backstep → short bias

   • hover → neutral

   • perch → risk-off (volatility extreme)

6. Trade only on a new closed bar (to avoid mid-bar churn). If there’s no new bar, you still see live probabilities and status on the HUD.

Inputs you can tune

• Core lengths: EMA short/long, RSI length, ATR length, Bollinger length/dev, ADX length, “burst” window.

• Feature weights: W_EMA , W_MACD , W_RSI , W_VWAP , W_BURST , W_SQZ , W_VOL . These scale each component’s influence.

• Decision shaping: SoftmaxTemp (how “sharp” probabilities are), ProbLongThresh , ProbShortThresh , CooldownBars .

• Volatility regime: ATRZLen (window for ATR% z-score) and ATRZThresh (threshold for triggering risk-off “perch”).

• Risk: Lots , and ATR-based ATRmultSL / ATRmultTP .

Indicators & features (how they’re built)

• EMA short/long trend: Trend score uses the normalized difference (EMA_short – EMA_long) / avg(EMA_short, EMA_long) .

• MACD momentum: Uses MACD_main – MACD_signal (the classic histogram) as a momentum proxy.

• RSI balance: ((RSI – 50) / 50) to center RSI around 0.

• Bollinger width & position: Width (upper–lower) normalized by the middle band; position within the band can be used for context; width contributes to the squeeze calculation.

• ATR & ATR%: ATR / close gives scale-free volatility. The EA also computes an ATR% z-score over ATRZLen prior bars to identify volatility extremes.

• “Perch” (risk-off): Triggered when the ATR% z-score exceeds ATRZThresh . In this state the EA closes any open position and stands down.

• Session VWAP: Rebuilt each day using typical price (H+L+C)/3 times volume, divided by cumulative volume; vwapSpread = (close – VWAP)/close gauges stretched conditions relative to fair value.

• Burst (micro-impulse): Sum of 1-bar returns over the last InpBurstLen bars; a quick way to capture short-horizon pressure.

• Squeeze: ATR% / BollingerWidth —high when volatility is large relative to the price’s envelope width.

From features → probabilities

Each feature is multiplied by its weight and combined into three raw scores:

• rawLong accentuates trend + momentum + favorable spread + burst + (inverse) squeeze and penalizes volatility extremes.

• rawShort mirrors rawLong in sign (favoring the bearish perspective).

• rawHover is a small constant prior (keeps neutral viable when evidence is thin).

These raw scores go through a softmax (with SoftmaxTemp ) to produce:

• pLong, pNeut, pShort ∈ [0, 1]; the HUD shows them as percent integers.

Decision logic (state machine)

• If perch (volatility extreme): immediately risk-off (close any position, take no new trades).

• Otherwise, if not in cooldown:

  • If pLong ≥ ProbLongThresh and pLong > pShort → dart (long).

  • Else if pShort ≥ ProbShortThresh and pShort > pLong → backstep (short).

  • Else → hover (neutral).

• Cooldown: After a dart or backstep, the EA waits CooldownBars bars before allowing a new decisive trade, to reduce whipsaw.

Trade execution & risk

• When trades fire: Only once per new closed bar (the bar time advances). This keeps execution deterministic and non-repainting.

• Netting behavior: If an opposite position exists when a new signal hits, the EA closes it and flips.

• Stops/targets: Both are ATR-based, scaled by ATRmultSL and ATRmultTP . This adapts risk to volatility.

• “Perch” close: If risk-off triggers, any open position is closed immediately.

HUD (what you see on the chart)

• Status line: Shows whether indicators/history are ready (e.g., “Waiting for indicators”, “Ready”, “No history”).

• State: One of hover/dart/backstep/perch. Also shows (bar: new|hold) to indicate whether the current evaluation coincides with a newly closed bar ( new ) or the last closed one ( hold ).

• Probabilities: pLong / pNeut / pShort as whole-number percentages.

• Context: atrpZ (volatility regime) and vwapSpread (stretch vs VWAP).

• The HUD updates continuously (using the latest closed bar), even if markets are closed or ticks are slow; a 1-second timer can also drive updates.

Resilience & attachment behavior

• Will attach even when markets are closed: If indicators can’t be created immediately or history is thin, the EA doesn’t abort. It displays status until things are ready, then automatically begins evaluating and (on new bars) trading.

• History guard: If there aren’t enough bars to compute the ATR z-score window plus features, it shows a friendly status and waits.

• Self-recovery: As soon as data/indicators become available, the HUD flips to “Ready” and the algorithm runs.

Practical notes

• Use the Strategy Tester (Visual) to verify behavior outside market hours.

• Symbols with exotic sessions: The daily VWAP reset uses calendar days. If your instrument trades on unusual sessions, you may want to customize the reset logic.

• Volume type: tick_volume is used (typical for MT5). On some brokers/symbols this is a proxy for real volume.

• Tuning: Markets differ. You’ll likely want to adjust the weights, probability thresholds, cooldown, and ATR z-score settings for your symbol/timeframe.

In short, the EA takes a “hummingbird” approach—fast, decisive, and conservative under stress—by converting a basket of familiar signals into clear, probabilistic calls, acting only at bar closes, and stepping aside in turbulent regimes.