Курнат Андрей
504 lines
18 KiB
Python
504 lines
18 KiB
Python
from __future__ import annotations
|
|
|
|
import json
|
|
import math
|
|
from dataclasses import asdict, dataclass, field
|
|
from functools import lru_cache
|
|
from typing import Any
|
|
|
|
from crypto_spot_bot.config import Settings
|
|
from crypto_spot_bot.models import Candle
|
|
|
|
|
|
@dataclass(slots=True)
|
|
class TimeSeriesForecast:
|
|
enabled: bool
|
|
usable: bool
|
|
model: str
|
|
volatility_model: str
|
|
expected_return_percent: float
|
|
expected_price: float
|
|
volatility_percent: float
|
|
probability_up: float
|
|
confidence_adjustment: float
|
|
block_entry: bool
|
|
validation_mae_percent: float
|
|
baseline_mae_percent: float
|
|
skill: float
|
|
horizon: int
|
|
reason: str
|
|
candidates: list[dict[str, Any]] = field(default_factory=list)
|
|
|
|
def as_dict(self) -> dict[str, Any]:
|
|
return asdict(self)
|
|
|
|
|
|
class TimeSeriesForecaster:
|
|
def __init__(self, settings: Settings):
|
|
self.settings = settings
|
|
self._lstm_artifact_mtime: float | None = None
|
|
self._lstm_artifact: dict[str, Any] = {}
|
|
|
|
def forecast(self, candles: list[Candle], symbol: str | None = None) -> TimeSeriesForecast:
|
|
if not self.settings.time_series_forecast_enabled:
|
|
return _empty_forecast(False, "прогноз временных рядов выключен")
|
|
closes = [float(candle.close) for candle in candles if candle.close > 0]
|
|
min_candles = max(30, self.settings.time_series_min_candles)
|
|
if len(closes) < min_candles:
|
|
return _empty_forecast(True, "недостаточно свечей для прогноза")
|
|
returns = _log_returns(closes)
|
|
if len(returns) < 20:
|
|
return _empty_forecast(True, "недостаточно доходностей для прогноза")
|
|
|
|
validation_window = min(
|
|
max(8, self.settings.time_series_validation_window),
|
|
max(8, len(returns) // 3),
|
|
)
|
|
lstm_artifact = self._load_lstm_artifact()
|
|
candidates = _validate_candidates(returns, validation_window, self.settings, symbol, lstm_artifact)
|
|
best = min(candidates, key=lambda item: item["mae"])
|
|
baseline = next(item for item in candidates if item["model"] == "naive")
|
|
latest_prediction = _predict_next_return(best["model"], returns, self.settings, symbol, lstm_artifact)
|
|
horizon = max(1, self.settings.time_series_forecast_horizon)
|
|
expected_return = latest_prediction * horizon
|
|
expected_price = closes[-1] * math.exp(expected_return)
|
|
|
|
ewma_vol = _ewma_volatility(returns, self.settings.time_series_ewma_lambda)
|
|
garch_vol = _fixed_garch_volatility(returns)
|
|
vol_one_step = max(ewma_vol, garch_vol)
|
|
volatility_percent = vol_one_step * math.sqrt(horizon) * 100
|
|
expected_return_percent = (math.exp(expected_return) - 1) * 100
|
|
probability_up = _normal_cdf(expected_return / max(vol_one_step * math.sqrt(horizon), 1e-9))
|
|
baseline_mae = float(baseline["mae"])
|
|
model_mae = float(best["mae"])
|
|
skill = (baseline_mae - model_mae) / baseline_mae if baseline_mae > 0 else 0.0
|
|
skill = _clamp(skill, -1.0, 1.0)
|
|
min_edge = max(0.0, self.settings.time_series_min_edge_percent)
|
|
usable_skill = skill > 0.02 and best["model"] != "naive"
|
|
confidence_adjustment = _confidence_adjustment(
|
|
expected_return_percent=expected_return_percent,
|
|
probability_up=probability_up,
|
|
skill=skill,
|
|
min_edge=min_edge,
|
|
max_adjustment=self.settings.time_series_max_adjustment,
|
|
usable_skill=usable_skill,
|
|
)
|
|
block_entry = bool(
|
|
usable_skill
|
|
and expected_return_percent <= -min_edge
|
|
and probability_up <= 0.45
|
|
)
|
|
reason = _reason(
|
|
model=best["model"],
|
|
expected_return_percent=expected_return_percent,
|
|
probability_up=probability_up,
|
|
skill=skill,
|
|
block_entry=block_entry,
|
|
usable_skill=usable_skill,
|
|
)
|
|
return TimeSeriesForecast(
|
|
enabled=True,
|
|
usable=True,
|
|
model=best["model"],
|
|
volatility_model="max(EWMA,GARCH-like)",
|
|
expected_return_percent=round(expected_return_percent, 4),
|
|
expected_price=round(expected_price, 8),
|
|
volatility_percent=round(volatility_percent, 4),
|
|
probability_up=round(probability_up, 4),
|
|
confidence_adjustment=round(confidence_adjustment, 4),
|
|
block_entry=block_entry,
|
|
validation_mae_percent=round(model_mae * 100, 4),
|
|
baseline_mae_percent=round(baseline_mae * 100, 4),
|
|
skill=round(skill, 4),
|
|
horizon=horizon,
|
|
reason=reason,
|
|
candidates=[
|
|
{"model": item["model"], "mae_percent": round(float(item["mae"]) * 100, 4)}
|
|
for item in sorted(candidates, key=lambda item: item["mae"])
|
|
],
|
|
)
|
|
|
|
def _load_lstm_artifact(self) -> dict[str, Any]:
|
|
if not self.settings.time_series_lstm_enabled:
|
|
return {}
|
|
path = self.settings.time_series_lstm_model_path
|
|
try:
|
|
stat = path.stat()
|
|
except OSError:
|
|
self._lstm_artifact_mtime = None
|
|
self._lstm_artifact = {}
|
|
return {}
|
|
if self._lstm_artifact_mtime == stat.st_mtime:
|
|
return self._lstm_artifact
|
|
try:
|
|
data = json.loads(path.read_text(encoding="utf-8"))
|
|
except (OSError, json.JSONDecodeError):
|
|
data = {}
|
|
self._lstm_artifact = data if isinstance(data, dict) else {}
|
|
self._lstm_artifact_mtime = stat.st_mtime
|
|
return self._lstm_artifact
|
|
|
|
|
|
def _empty_forecast(enabled: bool, reason: str) -> TimeSeriesForecast:
|
|
return TimeSeriesForecast(
|
|
enabled=enabled,
|
|
usable=False,
|
|
model="none",
|
|
volatility_model="none",
|
|
expected_return_percent=0.0,
|
|
expected_price=0.0,
|
|
volatility_percent=0.0,
|
|
probability_up=0.5,
|
|
confidence_adjustment=0.0,
|
|
block_entry=False,
|
|
validation_mae_percent=0.0,
|
|
baseline_mae_percent=0.0,
|
|
skill=0.0,
|
|
horizon=0,
|
|
reason=reason,
|
|
)
|
|
|
|
|
|
def _log_returns(closes: list[float]) -> list[float]:
|
|
return [math.log(closes[index] / closes[index - 1]) for index in range(1, len(closes))]
|
|
|
|
|
|
def _validate_candidates(
|
|
returns: list[float],
|
|
validation_window: int,
|
|
settings: Settings,
|
|
symbol: str | None = None,
|
|
lstm_artifact: dict[str, Any] | None = None,
|
|
) -> list[dict[str, float | str]]:
|
|
models = ["naive", "drift", "ewma", "ar1", "ar3"]
|
|
if _can_use_lstm(returns, settings, symbol, lstm_artifact or {}):
|
|
models.append("lstm")
|
|
rows: list[dict[str, float | str]] = []
|
|
start = max(8, len(returns) - validation_window)
|
|
for model in models:
|
|
errors: list[float] = []
|
|
for index in range(start, len(returns)):
|
|
history = returns[:index]
|
|
if len(history) < 8:
|
|
continue
|
|
predicted = _predict_next_return(model, history, settings, symbol, lstm_artifact)
|
|
errors.append(abs(predicted - returns[index]))
|
|
mae = sum(errors) / len(errors) if errors else 1e9
|
|
rows.append({"model": model, "mae": mae})
|
|
return rows
|
|
|
|
|
|
def _predict_next_return(
|
|
model: str,
|
|
returns: list[float],
|
|
settings: Settings | None = None,
|
|
symbol: str | None = None,
|
|
lstm_artifact: dict[str, Any] | None = None,
|
|
) -> float:
|
|
if model == "naive":
|
|
return 0.0
|
|
if model == "drift":
|
|
window = returns[-24:] if len(returns) >= 24 else returns
|
|
return sum(window) / len(window) if window else 0.0
|
|
if model == "ewma":
|
|
return _ewma_mean(returns, 0.82)
|
|
if model == "ar1":
|
|
return _ar_predict(returns, 1)
|
|
if model == "ar3":
|
|
return _ar_predict(returns, 3)
|
|
if model == "lstm" and settings is not None:
|
|
return _lstm_predict(returns, settings, symbol, lstm_artifact or {})
|
|
return 0.0
|
|
|
|
|
|
def _ewma_mean(values: list[float], decay: float) -> float:
|
|
if not values:
|
|
return 0.0
|
|
estimate = values[0]
|
|
alpha = 1 - _clamp(decay, 0.01, 0.99)
|
|
for value in values[1:]:
|
|
estimate = alpha * value + (1 - alpha) * estimate
|
|
return estimate
|
|
|
|
|
|
def _ar_predict(returns: list[float], lag_count: int) -> float:
|
|
if len(returns) <= lag_count + 6:
|
|
return _predict_next_return("drift", returns)
|
|
rows: list[list[float]] = []
|
|
targets: list[float] = []
|
|
for index in range(lag_count, len(returns)):
|
|
rows.append([1.0] + [returns[index - lag] for lag in range(1, lag_count + 1)])
|
|
targets.append(returns[index])
|
|
coeffs = _ols(rows, targets)
|
|
if not coeffs:
|
|
return _predict_next_return("drift", returns)
|
|
features = [1.0] + [returns[-lag] for lag in range(1, lag_count + 1)]
|
|
prediction = sum(coeff * feature for coeff, feature in zip(coeffs, features))
|
|
recent_abs = sorted(abs(value) for value in returns[-48:]) if len(returns) >= 8 else [0.01]
|
|
cap = max(recent_abs[int(len(recent_abs) * 0.9)], 0.0002)
|
|
return _clamp(prediction, -cap, cap)
|
|
|
|
|
|
def _can_use_lstm(
|
|
returns: list[float],
|
|
settings: Settings,
|
|
symbol: str | None,
|
|
lstm_artifact: dict[str, Any],
|
|
) -> bool:
|
|
if not settings.time_series_lstm_enabled:
|
|
return False
|
|
params = _lstm_params(settings, symbol, lstm_artifact)
|
|
return len(returns) >= params["lookback"] + 16
|
|
|
|
|
|
def _lstm_params(settings: Settings, symbol: str | None, lstm_artifact: dict[str, Any]) -> dict[str, float | int]:
|
|
params: dict[str, float | int] = {
|
|
"lookback": settings.time_series_lstm_lookback,
|
|
"units": settings.time_series_lstm_units,
|
|
"ridge": settings.time_series_lstm_ridge,
|
|
}
|
|
default_params = lstm_artifact.get("default")
|
|
if isinstance(default_params, dict):
|
|
params.update(_clean_lstm_params(default_params))
|
|
symbols = lstm_artifact.get("symbols")
|
|
symbol_params = symbols.get(symbol.upper()) if symbol and isinstance(symbols, dict) else None
|
|
if isinstance(symbol_params, dict):
|
|
params.update(_clean_lstm_params(symbol_params))
|
|
return {
|
|
"lookback": int(_clamp(float(params["lookback"]), 6.0, 128.0)),
|
|
"units": int(_clamp(float(params["units"]), 2.0, 16.0)),
|
|
"ridge": _clamp(float(params["ridge"]), 1e-8, 0.5),
|
|
}
|
|
|
|
|
|
def _clean_lstm_params(data: dict[str, Any]) -> dict[str, float | int]:
|
|
clean: dict[str, float | int] = {}
|
|
for key in ("lookback", "units", "ridge"):
|
|
value = data.get(key)
|
|
if isinstance(value, (int, float)):
|
|
clean[key] = value
|
|
elif isinstance(value, str):
|
|
try:
|
|
clean[key] = float(value)
|
|
except ValueError:
|
|
continue
|
|
return clean
|
|
|
|
|
|
def _lstm_predict(
|
|
returns: list[float],
|
|
settings: Settings,
|
|
symbol: str | None,
|
|
lstm_artifact: dict[str, Any],
|
|
) -> float:
|
|
params = _lstm_params(settings, symbol, lstm_artifact)
|
|
lookback = int(params["lookback"])
|
|
units = int(params["units"])
|
|
ridge = float(params["ridge"])
|
|
if len(returns) <= lookback + 8:
|
|
return _predict_next_return("drift", returns)
|
|
|
|
scale = _return_scale(returns)
|
|
normalized = [_clamp(value / scale, -6.0, 6.0) for value in returns]
|
|
states = _lstm_states(normalized, units)
|
|
rows: list[list[float]] = []
|
|
targets: list[float] = []
|
|
for index in range(lookback, len(returns)):
|
|
rows.append([1.0] + states[index - 1])
|
|
targets.append(normalized[index])
|
|
coeffs = _ols(rows, targets, ridge)
|
|
if not coeffs:
|
|
return _predict_next_return("drift", returns)
|
|
|
|
features = [1.0] + states[-1]
|
|
prediction = sum(coeff * feature for coeff, feature in zip(coeffs, features))
|
|
prediction = _clamp(prediction, -4.0, 4.0) * scale
|
|
recent_abs = sorted(abs(value) for value in returns[-48:]) if len(returns) >= 8 else [0.01]
|
|
cap = max(recent_abs[int(len(recent_abs) * 0.9)], 0.0002)
|
|
return _clamp(prediction, -cap, cap)
|
|
|
|
|
|
def _return_scale(returns: list[float]) -> float:
|
|
recent = returns[-120:] if len(returns) > 120 else returns
|
|
values = sorted(abs(value) for value in recent if math.isfinite(value))
|
|
if not values:
|
|
return 0.0005
|
|
median = values[len(values) // 2]
|
|
mean = sum(values) / len(values)
|
|
return max(max(median, mean * 0.5), 1e-5)
|
|
|
|
|
|
def _lstm_states(normalized_returns: list[float], units: int) -> list[list[float]]:
|
|
weights = _lstm_weights(units)
|
|
hidden = [0.0 for _ in range(units)]
|
|
cell = [0.0 for _ in range(units)]
|
|
states: list[list[float]] = []
|
|
for value in normalized_returns:
|
|
hidden, cell = _lstm_step(value, hidden, cell, weights)
|
|
states.append(hidden[:])
|
|
return states
|
|
|
|
|
|
@lru_cache(maxsize=16)
|
|
def _lstm_weights(units: int) -> tuple[list[list[float]], list[list[list[float]]], list[list[float]]]:
|
|
input_weights: list[list[float]] = []
|
|
recurrent_weights: list[list[list[float]]] = []
|
|
biases: list[list[float]] = []
|
|
base_biases = (-0.15, 0.70, 0.05, 0.0)
|
|
for gate in range(4):
|
|
gate_input: list[float] = []
|
|
gate_recurrent: list[list[float]] = []
|
|
gate_bias: list[float] = []
|
|
for unit in range(units):
|
|
gate_input.append(0.55 * math.sin((gate + 1) * (unit + 1) * 1.61803398875))
|
|
gate_recurrent.append(
|
|
[
|
|
0.14 * math.sin((gate + 3) * (unit + 1) * (source + 1) * 0.731)
|
|
for source in range(units)
|
|
]
|
|
)
|
|
gate_bias.append(base_biases[gate] + 0.03 * math.sin((gate + 1) * (unit + 1)))
|
|
input_weights.append(gate_input)
|
|
recurrent_weights.append(gate_recurrent)
|
|
biases.append(gate_bias)
|
|
return input_weights, recurrent_weights, biases
|
|
|
|
|
|
def _lstm_step(
|
|
value: float,
|
|
hidden: list[float],
|
|
cell: list[float],
|
|
weights: tuple[list[list[float]], list[list[list[float]]], list[list[float]]],
|
|
) -> tuple[list[float], list[float]]:
|
|
input_weights, recurrent_weights, biases = weights
|
|
units = len(hidden)
|
|
next_hidden = [0.0 for _ in range(units)]
|
|
next_cell = [0.0 for _ in range(units)]
|
|
for unit in range(units):
|
|
gate_values = []
|
|
for gate in range(4):
|
|
raw = input_weights[gate][unit] * value + biases[gate][unit]
|
|
raw += sum(recurrent_weights[gate][unit][source] * hidden[source] for source in range(units))
|
|
gate_values.append(raw)
|
|
input_gate = _sigmoid(gate_values[0])
|
|
forget_gate = _sigmoid(gate_values[1])
|
|
output_gate = _sigmoid(gate_values[2])
|
|
candidate = math.tanh(gate_values[3])
|
|
next_cell[unit] = forget_gate * cell[unit] + input_gate * candidate
|
|
next_hidden[unit] = output_gate * math.tanh(next_cell[unit])
|
|
return next_hidden, next_cell
|
|
|
|
|
|
def _sigmoid(value: float) -> float:
|
|
if value >= 40:
|
|
return 1.0
|
|
if value <= -40:
|
|
return 0.0
|
|
return 1 / (1 + math.exp(-value))
|
|
|
|
|
|
def _ols(rows: list[list[float]], targets: list[float], ridge: float = 1e-8) -> list[float] | None:
|
|
if not rows:
|
|
return None
|
|
columns = len(rows[0])
|
|
xtx = [[0.0 for _ in range(columns)] for _ in range(columns)]
|
|
xty = [0.0 for _ in range(columns)]
|
|
for row, target in zip(rows, targets):
|
|
for i in range(columns):
|
|
xty[i] += row[i] * target
|
|
for j in range(columns):
|
|
xtx[i][j] += row[i] * row[j]
|
|
for i in range(columns):
|
|
xtx[i][i] += ridge
|
|
return _solve_linear_system(xtx, xty)
|
|
|
|
|
|
def _solve_linear_system(matrix: list[list[float]], vector: list[float]) -> list[float] | None:
|
|
size = len(vector)
|
|
augmented = [row[:] + [vector[index]] for index, row in enumerate(matrix)]
|
|
for col in range(size):
|
|
pivot = max(range(col, size), key=lambda row: abs(augmented[row][col]))
|
|
if abs(augmented[pivot][col]) < 1e-12:
|
|
return None
|
|
augmented[col], augmented[pivot] = augmented[pivot], augmented[col]
|
|
pivot_value = augmented[col][col]
|
|
for item in range(col, size + 1):
|
|
augmented[col][item] /= pivot_value
|
|
for row in range(size):
|
|
if row == col:
|
|
continue
|
|
factor = augmented[row][col]
|
|
for item in range(col, size + 1):
|
|
augmented[row][item] -= factor * augmented[col][item]
|
|
return [augmented[row][size] for row in range(size)]
|
|
|
|
|
|
def _ewma_volatility(returns: list[float], decay: float) -> float:
|
|
if not returns:
|
|
return 0.0
|
|
decay = _clamp(decay, 0.80, 0.995)
|
|
variance = returns[0] * returns[0]
|
|
for value in returns[1:]:
|
|
variance = decay * variance + (1 - decay) * value * value
|
|
return math.sqrt(max(variance, 0.0))
|
|
|
|
|
|
def _fixed_garch_volatility(returns: list[float]) -> float:
|
|
if not returns:
|
|
return 0.0
|
|
long_variance = sum(value * value for value in returns) / len(returns)
|
|
alpha = 0.08
|
|
beta = 0.90
|
|
omega = max(1e-12, (1 - alpha - beta) * long_variance)
|
|
variance = long_variance
|
|
for value in returns:
|
|
variance = omega + alpha * value * value + beta * variance
|
|
return math.sqrt(max(variance, 0.0))
|
|
|
|
|
|
def _confidence_adjustment(
|
|
*,
|
|
expected_return_percent: float,
|
|
probability_up: float,
|
|
skill: float,
|
|
min_edge: float,
|
|
max_adjustment: float,
|
|
usable_skill: bool,
|
|
) -> float:
|
|
if not usable_skill:
|
|
return 0.0
|
|
edge = abs(expected_return_percent) - min_edge
|
|
if edge <= 0:
|
|
return 0.0
|
|
direction = 1.0 if expected_return_percent > 0 and probability_up >= 0.55 else -1.0
|
|
if direction < 0 and probability_up > 0.45:
|
|
return 0.0
|
|
strength = _clamp(edge / max(min_edge, 0.05), 0.0, 1.0)
|
|
probability_strength = _clamp(abs(probability_up - 0.5) / 0.25, 0.0, 1.0)
|
|
skill_strength = _clamp(skill / 0.18, 0.0, 1.0)
|
|
return direction * _clamp(max_adjustment, 0.0, 0.18) * strength * probability_strength * skill_strength
|
|
|
|
|
|
def _reason(
|
|
*,
|
|
model: str,
|
|
expected_return_percent: float,
|
|
probability_up: float,
|
|
skill: float,
|
|
block_entry: bool,
|
|
usable_skill: bool,
|
|
) -> str:
|
|
if not usable_skill:
|
|
return f"модель {model} не лучше baseline на walk-forward проверке"
|
|
if block_entry:
|
|
return f"модель {model}: ожидаемое движение вниз {expected_return_percent:.3f}%, P(рост)={probability_up:.2f}"
|
|
return f"модель {model}: прогноз {expected_return_percent:.3f}%, P(рост)={probability_up:.2f}, skill={skill:.3f}"
|
|
|
|
|
|
def _normal_cdf(value: float) -> float:
|
|
return 0.5 * (1 + math.erf(value / math.sqrt(2)))
|
|
|
|
|
|
def _clamp(value: float, low: float, high: float) -> float:
|
|
return max(low, min(high, value))
|