The Feature Engine

🦉 AthenaAI is an AI-powered ensemble machine-learning trading bot for PocketOption, built on BinaryOptionsToolsV2 → Get it on GitLab
💬 Need help? Join the Chipa Discord · ⚡ Go no-code with ChipaEditor

FeatureEngine (in src/core/feature_engine.py) converts a window of candles into the numeric feature vector that feeds the ensemble predictor, the expiry selector, and the indicator-alignment gate in src/bot.py. The vector always has 57 slots: 40 core features (indices 0–39, always active) followed by 17 experimental features (indices 40–56, individually maskable by the Feature Lab).

compute(candles, window=20) takes the bot’s candle history (up to BotConfig.lookback, default 200 candles) and the rolling window (BotConfig.feature_window, default 20). All indicators are implemented in-house as static helpers on the class — there is no TA-lib dependency.

The ≥26-candle minimum

if len(candles) < max(window, 26):
    return None

compute returns None until at least max(window, 26) candles exist — 26 because the MACD’s slow EMA needs 26 closes. The trade loop treats a None result as a skip (log line ⏸ Feature compute returned None). Separately, src/bot.py enforces an effective warmup of at least 60 candles before trading at all, since experimental features such as sma50_dist need 50 candles to produce non-zero values (features that lack enough history return neutral defaults rather than failing).

NaN handling

The engine itself guards divisions with + 1e-10 epsilons, but callers additionally sanitize the output before use:

features = np.nan_to_num(features, nan=0.0, posinf=0.0, neginf=0.0)

This happens both in the live trade loop and in dataset pre-training (src/bot.py), so a pathological candle window can never inject NaN/Inf into the models.

Core features (indices 0–39)

Prices below are raw quote prices (for EURUSD, roughly 1.0–1.2), so absolute-price-difference features are on the order of 0.0001 (one pip = 0.0001).

Price action

Index	Name	Formula / description	Range / typical
0	`body`	`close[-1] - open[-1]` — signed body of the current candle	± a few pips (±0.0005)
1	`range`	`high[-1] - low[-1]` — current candle’s full range	≥ 0, typically 0.0001–0.001
2	`mom_1`	`close[-1] - close[-2]` — 1-bar momentum	± a few pips
3	`mom_5`	`close[-1] - close[-5]` (0 if fewer than 5 closes)	± tens of pips

Returns

rets = diff(closes) / closes[:-1] (per-bar fractional returns).

Index	Name	Formula / description	Range / typical
4	`ret_last`	Latest return `rets[-1]`	≈ ±0.001 on 1-min FX
5	`ret_mean`	`mean(rets[-window:])` — mean return over the window	≈ ±0.0002
6	`volatility`	`std(rets[-window:])` — return standard deviation	≥ 0, ~0.0001–0.001 (also read by the expiry selector)

Moving averages and MACD

Index	Name	Formula / description	Range / typical
7	`price_sma5`	`close[-1] - SMA(5)`	± pips
8	`price_sma10`	`close[-1] - SMA(10)`	± pips
9	`price_sma20`	`close[-1] - SMA(window)`	± pips
10	`ma_cross`	`SMA(5) - SMA(20)` — fast/slow crossover; sign = short-term trend. Used by the indicator-alignment gate and the fallback predictor	± pips
11	`macd_line`	`EMA(12) - EMA(26)` (scalar EMAs of closes)	± pips
12	`macd_hist`	MACD histogram: `macd_line_array[-1] - EMA(macd_line_array, 9)`. Used by the indicator-alignment gate	± fractions of a pip

Oscillators

Index	Name	Formula / description	Range / typical
13	`rsi`	14-period RSI (simple-mean gains/losses over the last 15 closes); 50.0 if insufficient data, 100.0 if zero losses	0–100
14	`rsi_zone`	`+1.0` if RSI > 70, `−1.0` if RSI < 30, else `0.0`	{−1, 0, +1}
15	`bb_width`	Bollinger width `(upper − lower) / mid` with `mid = SMA(window)`, bands at ±2 std of closes	≥ 0, ~0.0005–0.005
16	`bb_pos`	Position within bands `(close − lower) / (upper − lower)`	~0–1 (can slightly overshoot)
17	`atr`	14-period ATR (mean true range)	≥ 0, ~0.0001–0.001 (read by expiry selector)
18	`range_vs_atr`	`(high[-1] − low[-1]) / atr` — is the current candle unusually large?	≥ 0, ~0.5–3
19	`stoch_k`	Stochastic %K over 14 periods: `100·(close − low_min)/(high_max − low_min)`	0–100
20	`stoch_d`	%D — mean of the last 3 %K values	0–100
21	`stoch_diff`	`%K − %D`	≈ ±100, usually ±20
22	`cci`	20-period Commodity Channel Index on typical price `(H+L+C)/3`, scaled by `0.015 · MAD`	unbounded, usually ±200
23	`williams_r`	14-period Williams %R: `−100·(HH − close)/(HH − LL)`	−100 to 0
24	`adx`	Simplified ADX: DX from smoothed +DM/−DM over ATR — a single DX reading, not a smoothed ADX average	0–100 (read by expiry selector)

Volume

If all volumes are zero (common for FX quote feeds), these default to 1.0 and 0.0.

Index	Name	Formula / description	Range / typical
25	`rel_volume`	`vol[-1] / mean(vol[-window:])`	≥ 0, ~1.0
26	`vol_std`	`std(vol[-window:]) / mean(vol[-window:])` (coefficient of variation)	≥ 0

Candle patterns

Index	Name	Formula / description	Range / typical
27	`doji`	`1.0` if body/range < 0.1 (indecision candle), else `0.0`	{0, 1}
28	`hammer`	`+1.0` hammer (lower wick > 2× body, small upper wick); `−1.0` inverted hammer; else `0.0`	{−1, 0, +1}
29	`engulfing`	`+1.0` bullish engulfing (bear candle followed by larger bull body); `−1.0` bearish engulfing; else `0.0`	{−1, 0, +1}

Higher-order statistics

Computed on rets[-window:].

Index	Name	Formula / description	Range / typical
30	`ret_p25`	25th percentile of windowed returns	≤ 0 usually
31	`ret_p75`	75th percentile of windowed returns	≥ 0 usually
32	`skew`	Sample skewness `mean(((r−μ)/σ)³)`	unbounded, usually ±2
33	`kurt`	Excess kurtosis `mean(((r−μ)/σ)⁴) − 3`	unbounded, usually −2 to +10
34	`fractal_dim`	Higuchi fractal dimension of the last `window` closes (kmax=5, log-log slope); 1.5 fallback when the series is too short or degenerate	~1.0 (trending) to ~2.0 (noisy)

Streak and time

Index	Name	Formula / description	Range / typical
35	`streak`	Signed count of consecutive up (+) or down (−) closes ending at the latest candle, capped at ±10, terminated by a flat close	−10 to +10
36	`time_sin_h`	`sin(2π · hour/24)` with fractional hour, UTC	−1 to 1
37	`time_cos_h`	`cos(2π · hour/24)`	−1 to 1
38	`time_sin_d`	`sin(2π · weekday/7)` (Monday = 0)	−1 to 1
39	`time_cos_d`	`cos(2π · weekday/7)`	−1 to 1

The sine/cosine pairs encode time cyclically so 23:59 and 00:01 are numerically adjacent.

Experimental features (indices 40–56)

These sit behind the experimental_enabled master switch (default True; when off, all 17 slots are zero-filled to keep the vector length stable) and are the only features the Feature Lab may mask.

Candlestick anatomy

Index	Name	Formula / description	Range / typical
40	`shooting_star`	`1.0` if upper wick > 2× body and lower wick < body (and body > 0), else `0.0`	{0, 1}
41	`bearish_engulfing`	`1.0` if previous candle bullish, current bearish, and current body engulfs previous (`open[-1] > close[-2]` and `close[-1] < open[-2]`)	{0, 1}
55	`candle_wick_ratio`	`(range − body) / body` — total wick length relative to body	≥ 0, large for dojis
56	`body_vs_avg`	current `	body

Long trend

Index	Name	Formula / description	Range / typical
42	`sma50_dist`	`close[-1] − SMA(50)`; `0.0` if fewer than 50 closes	± tens of pips
43	`sma50_above`	`+1.0` if close above SMA(50), else `−1.0`; `0.0` if insufficient data	{−1, 0, +1}
44	`momentum_ratio`	`SMA(14) / SMA(28)`; `1.0` if fewer than 28 closes	~1.0 ± 0.001

Fibonacci retracement distances

Over the last 50 candles’ high/low range (all 0.0 if fewer than 50 candles). Each is the signed distance of the current price from the retracement level, normalized by the range:

Index	Name	Formula / description	Range / typical
45	`fib_dist_236`	`(price − (high − 0.236·range)) / range`	~−1 to +1
46	`fib_dist_382`	`(price − (high − 0.382·range)) / range`	~−1 to +1
47	`fib_dist_500`	`(price − (high − 0.500·range)) / range`	~−1 to +1
48	`fib_dist_618`	`(price − (high − 0.618·range)) / range`	~−1 to +1

Support / resistance and ratios

Index	Name	Formula / description	Range / typical
49	`support_dist`	`(price − min(lows[-window:])) / support` — fractional distance above windowed low	≥ 0, tiny (~0.0005)
50	`resist_dist`	`(max(highs[-window:]) − price) / price` — fractional distance below windowed high	≥ 0, tiny
51	`hl_ratio`	`high[-1] / low[-1]`	~1.0 + a few ×10⁻⁴
52	`oc_ratio`	`open[-1] / close[-1]`	~1.0 ± a few ×10⁻⁴

Divergence and volume spike

Index	Name	Formula / description	Range / typical
53	`rsi_divergence`	`+1.0` if price rose over 14 bars while RSI fell by more than 5 points (bearish divergence); `−1.0` for the mirror case; else `0.0`. Requires ≥28 closes (recomputes RSI at each of the last 14 bars)	{−1, 0, +1}
54	`volume_spike`	`vol[-1] / mean(vol[-window:])`; `1.0` when volume data is absent — numerically the same as `rel_volume` (index 25)	≥ 0, ~1.0

The mask mechanism

The engine holds a feature_mask — a length-57 float array of 1.0s at construction. Just before returning, compute multiplies element-wise:

result = np.array(feats, dtype=np.float64)
if len(result) == len(self.feature_mask):
    result *= self.feature_mask
return result

A masked feature (mask value 0.0) is therefore zeroed, not removed — the vector length never changes, so the ensemble’s scaler and models keep their dimensions and no reset is triggered. Only the Feature Lab writes to the mask, and only for experimental indices (40+); core features are always active. The mask is in-memory only — it resets to all-ones on restart and is rebuilt as the Feature Lab accumulates trade outcomes again.