Add PyTorch recurrent forecaster

README.md

@@ -72,6 +72,22 @@ python tools\train_lstm_forecaster.py --symbols BTCUSDT,ETHUSDT,SOLUSDT,XRPUSDT,
 
 Файл из `TIME_SERIES_LSTM_MODEL_PATH` читается ботом автоматически. Даже если LSTM-параметры сохранены, сделка меняется только тогда, когда текущая walk-forward проверка в `crypto_spot_bot/time_series.py` показывает качество лучше baseline.
 
+Для более тяжелого локального обучения можно использовать настоящий PyTorch `LSTM/GRU` trainer. PyTorch нужен только на машине обучения; в JSON экспортируются веса, а runtime на Raspberry Pi считает inference обычным Python-кодом:
+
+```powershell
+.\.venv\Scripts\python.exe -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+.\.venv\Scripts\python.exe tools\train_torch_recurrent_forecaster.py `
+  --symbols BTCUSDT,ETHUSDT,SOLUSDT,XRPUSDT,LTCUSDT `
+  --limit 1000 `
+  --architectures lstm,gru `
+  --lookbacks 32,64 `
+  --hidden-sizes 16,32 `
+  --layers 1 `
+  --epochs 60
+```
+
+Экспортированные модели появляются в dashboard как `torch_lstm` или `torch_gru`; легкий `lstm`-кандидат остается доступен как fallback.
+
 Автопереобучение запускает тот же train-скрипт, пишет лог в `runtime/lstm_retrain.log` и защищается от параллельных запусков:
 
 ```powershell
@@ -86,7 +102,7 @@ bash tools/run_lstm_retrain.sh
 bash tools/install_lstm_retrainer_systemd.sh
 ```
 
-По умолчанию расписание переобучает каждые 6 часов с `--limit 1000`; Windows-установщик фиксирует пары `BTCUSDT,ETHUSDT,SOLUSDT,XRPUSDT,LTCUSDT`, чтобы первый scheduled run был предсказуемым. Параметры можно переопределить через env: `LSTM_RETRAIN_SYMBOLS`, `LSTM_RETRAIN_LIMIT`, `LSTM_RETRAIN_LOOKBACKS`, `LSTM_RETRAIN_UNITS`, `LSTM_RETRAIN_RIDGES`, `LSTM_RETRAIN_INTERVAL`, `LSTM_RETRAIN_ENV`.
+По умолчанию Windows-расписание переобучает PyTorch `LSTM/GRU` каждые 6 часов с `--limit 1000`; Windows-установщик фиксирует пары `BTCUSDT,ETHUSDT,SOLUSDT,XRPUSDT,LTCUSDT`, чтобы первый scheduled run был предсказуемым. Параметры можно переопределить через env: `LSTM_RETRAIN_SYMBOLS`, `LSTM_RETRAIN_LIMIT`, `LSTM_RETRAIN_LOOKBACKS`, `LSTM_RETRAIN_ARCHITECTURES`, `LSTM_RETRAIN_HIDDEN_SIZES`, `LSTM_RETRAIN_LAYERS`, `LSTM_RETRAIN_DROPOUTS`, `LSTM_RETRAIN_EPOCHS`, `LSTM_RETRAIN_PATIENCE`, `LSTM_RETRAIN_INTERVAL`, `LSTM_RETRAIN_ENV`. Для старого легкого trainer можно запустить `tools\run_lstm_retrain.ps1 -Trainer reservoir`.
 
 ## Docker
 

crypto_spot_bot/time_series.py

@@ -171,6 +171,9 @@ def _validate_candidates(
     lstm_artifact: dict[str, Any] | None = None,
 ) -> list[dict[str, float | str]]:
     models = ["naive", "drift", "ewma", "ar1", "ar3"]
+    torch_model = _torch_recurrent_model_name(symbol, lstm_artifact or {})
+    if torch_model and _can_use_torch_recurrent(returns, symbol, lstm_artifact or {}):
+        models.append(torch_model)
     if _can_use_lstm(returns, settings, symbol, lstm_artifact or {}):
         models.append("lstm")
     rows: list[dict[str, float | str]] = []
@@ -206,6 +209,8 @@ def _predict_next_return(
         return _ar_predict(returns, 1)
     if model == "ar3":
         return _ar_predict(returns, 3)
+    if model in {"torch_lstm", "torch_gru"}:
+        return _torch_recurrent_predict(returns, symbol, lstm_artifact or {})
     if model == "lstm" and settings is not None:
         return _lstm_predict(returns, settings, symbol, lstm_artifact or {})
     return 0.0
@@ -285,6 +290,235 @@ def _clean_lstm_params(data: dict[str, Any]) -> dict[str, float | int]:
     return clean
 
 
+def _torch_recurrent_model_name(symbol: str | None, lstm_artifact: dict[str, Any]) -> str | None:
+    entry = _torch_recurrent_entry(symbol, lstm_artifact)
+    if not entry:
+        return None
+    architecture = str(entry.get("architecture", "")).strip().lower()
+    if architecture in {"lstm", "gru"}:
+        return f"torch_{architecture}"
+    model = str(entry.get("model", "")).strip().lower()
+    return model if model in {"torch_lstm", "torch_gru"} else None
+
+
+def _torch_recurrent_entry(symbol: str | None, lstm_artifact: dict[str, Any]) -> dict[str, Any] | None:
+    symbols = lstm_artifact.get("symbols")
+    entry = symbols.get(symbol.upper()) if symbol and isinstance(symbols, dict) else None
+    if not isinstance(entry, dict):
+        default = lstm_artifact.get("default")
+        entry = default if isinstance(default, dict) else None
+    if not isinstance(entry, dict):
+        return None
+    if not isinstance(entry.get("state_dict"), dict):
+        return None
+    return entry
+
+
+def _can_use_torch_recurrent(returns: list[float], symbol: str | None, lstm_artifact: dict[str, Any]) -> bool:
+    entry = _torch_recurrent_entry(symbol, lstm_artifact)
+    if not entry:
+        return False
+    lookback = int(_clamp(_float_entry(entry, "lookback", 0.0), 4.0, 512.0))
+    hidden_size = int(_clamp(_float_entry(entry, "hidden_size", 0.0), 1.0, 512.0))
+    num_layers = int(_clamp(_float_entry(entry, "num_layers", 1.0), 1.0, 8.0))
+    return len(returns) >= lookback + 1 and hidden_size > 0 and num_layers > 0
+
+
+def _torch_recurrent_predict(
+    returns: list[float],
+    symbol: str | None,
+    lstm_artifact: dict[str, Any],
+) -> float:
+    entry = _torch_recurrent_entry(symbol, lstm_artifact)
+    model_name = _torch_recurrent_model_name(symbol, lstm_artifact)
+    if not entry or not model_name:
+        return _predict_next_return("drift", returns)
+    lookback = int(_clamp(_float_entry(entry, "lookback", 0.0), 4.0, 512.0))
+    hidden_size = int(_clamp(_float_entry(entry, "hidden_size", 0.0), 1.0, 512.0))
+    num_layers = int(_clamp(_float_entry(entry, "num_layers", 1.0), 1.0, 8.0))
+    mean = _float_entry(entry, "mean", 0.0)
+    scale = max(_float_entry(entry, "scale", _return_scale(returns)), 1e-8)
+    clip = _clamp(_float_entry(entry, "clip", 8.0), 1.0, 50.0)
+    if len(returns) < lookback:
+        return _predict_next_return("drift", returns)
+
+    normalized = [_clamp((value - mean) / scale, -clip, clip) for value in returns[-lookback:]]
+    try:
+        hidden = _torch_recurrent_hidden(
+            normalized,
+            entry=entry,
+            model_name=model_name,
+            hidden_size=hidden_size,
+            num_layers=num_layers,
+        )
+        if hidden is None:
+            return _predict_next_return("drift", returns)
+        head_weight = _float_vector(entry.get("head_weight"))
+        head_bias = _float_entry(entry, "head_bias", 0.0)
+        if len(head_weight) != hidden_size:
+            return _predict_next_return("drift", returns)
+        normalized_prediction = sum(weight * value for weight, value in zip(head_weight, hidden)) + head_bias
+        if not math.isfinite(normalized_prediction):
+            return _predict_next_return("drift", returns)
+        prediction = _clamp(normalized_prediction, -clip, clip) * scale + mean
+    except (IndexError, KeyError, TypeError, ValueError, OverflowError):
+        return _predict_next_return("drift", returns)
+
+    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 _torch_recurrent_hidden(
+    normalized: list[float],
+    *,
+    entry: dict[str, Any],
+    model_name: str,
+    hidden_size: int,
+    num_layers: int,
+) -> list[float] | None:
+    state = entry.get("state_dict")
+    if not isinstance(state, dict):
+        return None
+    h_layers = [[0.0 for _ in range(hidden_size)] for _ in range(num_layers)]
+    c_layers = [[0.0 for _ in range(hidden_size)] for _ in range(num_layers)]
+    for value in normalized:
+        layer_input = [value]
+        for layer in range(num_layers):
+            if model_name == "torch_lstm":
+                next_hidden, next_cell = _torch_lstm_step(layer_input, h_layers[layer], c_layers[layer], state, layer)
+                h_layers[layer] = next_hidden
+                c_layers[layer] = next_cell
+            elif model_name == "torch_gru":
+                h_layers[layer] = _torch_gru_step(layer_input, h_layers[layer], state, layer)
+            else:
+                return None
+            layer_input = h_layers[layer]
+    return h_layers[-1]
+
+
+def _torch_lstm_step(
+    inputs: list[float],
+    hidden: list[float],
+    cell: list[float],
+    state: dict[str, Any],
+    layer: int,
+) -> tuple[list[float], list[float]]:
+    hidden_size = len(hidden)
+    gates = _torch_gate_values(inputs, hidden, state, layer, gate_count=4)
+    input_gate = [_sigmoid(value) for value in gates[0]]
+    forget_gate = [_sigmoid(value) for value in gates[1]]
+    cell_gate = [math.tanh(value) for value in gates[2]]
+    output_gate = [_sigmoid(value) for value in gates[3]]
+    next_cell = [
+        forget_gate[index] * cell[index] + input_gate[index] * cell_gate[index]
+        for index in range(hidden_size)
+    ]
+    next_hidden = [
+        output_gate[index] * math.tanh(next_cell[index])
+        for index in range(hidden_size)
+    ]
+    return next_hidden, next_cell
+
+
+def _torch_gru_step(
+    inputs: list[float],
+    hidden: list[float],
+    state: dict[str, Any],
+    layer: int,
+) -> list[float]:
+    hidden_size = len(hidden)
+    weight_ih = _float_matrix(state[f"weight_ih_l{layer}"])
+    weight_hh = _float_matrix(state[f"weight_hh_l{layer}"])
+    bias_ih = _float_vector(state[f"bias_ih_l{layer}"])
+    bias_hh = _float_vector(state[f"bias_hh_l{layer}"])
+
+    def gate_input(gate: int) -> list[float]:
+        start = gate * hidden_size
+        output = []
+        for index in range(hidden_size):
+            row = start + index
+            output.append(_dot(weight_ih[row], inputs) + bias_ih[row])
+        return output
+
+    def gate_hidden(gate: int) -> list[float]:
+        start = gate * hidden_size
+        output = []
+        for index in range(hidden_size):
+            row = start + index
+            output.append(_dot(weight_hh[row], hidden) + bias_hh[row])
+        return output
+
+    reset_input = gate_input(0)
+    update_input = gate_input(1)
+    new_input = gate_input(2)
+    reset_hidden = gate_hidden(0)
+    update_hidden = gate_hidden(1)
+    new_hidden = gate_hidden(2)
+
+    reset_gate = [_sigmoid(reset_input[index] + reset_hidden[index]) for index in range(hidden_size)]
+    update_gate = [_sigmoid(update_input[index] + update_hidden[index]) for index in range(hidden_size)]
+    candidate = [
+        math.tanh(new_input[index] + reset_gate[index] * new_hidden[index])
+        for index in range(hidden_size)
+    ]
+    return [
+        (1 - update_gate[index]) * candidate[index] + update_gate[index] * hidden[index]
+        for index in range(hidden_size)
+    ]
+
+
+def _torch_gate_values(
+    inputs: list[float],
+    hidden: list[float],
+    state: dict[str, Any],
+    layer: int,
+    gate_count: int,
+) -> list[list[float]]:
+    hidden_size = len(hidden)
+    weight_ih = _float_matrix(state[f"weight_ih_l{layer}"])
+    weight_hh = _float_matrix(state[f"weight_hh_l{layer}"])
+    bias_ih = _float_vector(state[f"bias_ih_l{layer}"])
+    bias_hh = _float_vector(state[f"bias_hh_l{layer}"])
+    gates: list[list[float]] = []
+    for gate in range(gate_count):
+        values = []
+        start = gate * hidden_size
+        for index in range(hidden_size):
+            row = start + index
+            values.append(_dot(weight_ih[row], inputs) + _dot(weight_hh[row], hidden) + bias_ih[row] + bias_hh[row])
+        gates.append(values)
+    return gates
+
+
+def _float_entry(data: dict[str, Any], key: str, default: float) -> float:
+    value = data.get(key)
+    if isinstance(value, (int, float)):
+        return float(value)
+    if isinstance(value, str):
+        try:
+            return float(value)
+        except ValueError:
+            return default
+    return default
+
+
+def _float_vector(data: Any) -> list[float]:
+    if not isinstance(data, list):
+        return []
+    return [float(value) for value in data]
+
+
+def _float_matrix(data: Any) -> list[list[float]]:
+    if not isinstance(data, list):
+        return []
+    return [_float_vector(row) for row in data]
+
+
+def _dot(left: list[float], right: list[float]) -> float:
+    return sum(left[index] * right[index] for index in range(min(len(left), len(right))))
+
+
 def _lstm_predict(
     returns: list[float],
     settings: Settings,

requirements-train.txt

@@ -0,0 +1,5 @@
+# Optional local-only training dependency for tools/train_torch_recurrent_forecaster.py.
+# Prefer the CPU wheel on Windows:
+#   python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+torch>=2.5
+numpy>=2.0

tests/test_time_series.py

@@ -122,3 +122,50 @@ def test_time_series_forecaster_reads_lstm_artifact(make_settings, tmp_path) ->
 
     assert forecast.usable is True
     assert any(candidate["model"] == "lstm" for candidate in forecast.candidates)
+
+
+def test_time_series_forecaster_reads_torch_gru_artifact(make_settings, tmp_path) -> None:
+    artifact_path = tmp_path / "lstm_forecaster.json"
+    hidden_size = 2
+    artifact_path.write_text(
+        json.dumps(
+            {
+                "version": 2,
+                "type": "pytorch_recurrent_forecaster",
+                "symbols": {
+                    "BTCUSDT": {
+                        "model": "torch_gru",
+                        "architecture": "gru",
+                        "lookback": 8,
+                        "hidden_size": hidden_size,
+                        "num_layers": 1,
+                        "mean": 0.0,
+                        "scale": 0.001,
+                        "clip": 8.0,
+                        "state_dict": {
+                            "weight_ih_l0": [[0.0] for _ in range(3 * hidden_size)],
+                            "weight_hh_l0": [[0.0, 0.0] for _ in range(3 * hidden_size)],
+                            "bias_ih_l0": [0.0 for _ in range(3 * hidden_size)],
+                            "bias_hh_l0": [0.0 for _ in range(3 * hidden_size)],
+                        },
+                        "head_weight": [0.0, 0.0],
+                        "head_bias": 0.2,
+                    },
+                },
+            }
+        ),
+        encoding="utf-8",
+    )
+    settings = make_settings(
+        tmp_path,
+        time_series_min_candles=80,
+        time_series_validation_window=20,
+        time_series_lstm_enabled=True,
+        time_series_lstm_model_path=artifact_path,
+    )
+    returns = [0.00015 if index % 4 else -0.00005 for index in range(140)]
+
+    forecast = TimeSeriesForecaster(settings).forecast(_candles_from_returns(returns), symbol="BTCUSDT")
+
+    assert forecast.usable is True
+    assert any(candidate["model"] == "torch_gru" for candidate in forecast.candidates)

tools/install_windows_lstm_retrainer.ps1

@@ -3,7 +3,10 @@ param(
     [string]$TaskName = "TradeBot LSTM Retrainer",
     [int]$EveryHours = 6,
     [string]$Symbols = "BTCUSDT,ETHUSDT,SOLUSDT,XRPUSDT,LTCUSDT",
-    [int]$Limit = 1000
+    [int]$Limit = 1000,
+    [ValidateSet("torch", "reservoir")]
+    [string]$Trainer = "torch",
+    [int]$FirstRunMinutes = 0
 )
 
 $ErrorActionPreference = "Stop"
@@ -14,7 +17,7 @@ if (-not (Test-Path $Runner)) {
     throw "Runner not found: $Runner"
 }
 
-$actionArgs = "-NoProfile -ExecutionPolicy Bypass -File `"$Runner`""
+$actionArgs = "-NoProfile -ExecutionPolicy Bypass -File `"$Runner`" -Trainer $Trainer"
 if ($Symbols) {
     $actionArgs += " -Symbols `"$Symbols`""
 }
@@ -24,7 +27,7 @@ if ($Limit -gt 0) {
 $action = New-ScheduledTaskAction -Execute "powershell.exe" -Argument $actionArgs -WorkingDirectory $RepoRoot
 $trigger = New-ScheduledTaskTrigger `
     -Once `
-    -At (Get-Date).AddMinutes(5) `
+    -At (Get-Date).AddMinutes($(if ($FirstRunMinutes -gt 0) { $FirstRunMinutes } else { $EveryHours * 60 })) `
     -RepetitionInterval (New-TimeSpan -Hours $EveryHours) `
     -RepetitionDuration (New-TimeSpan -Days 3650)
 $principal = New-ScheduledTaskPrincipal `

tools/run_lstm_retrain.ps1

@@ -1,10 +1,18 @@
 [CmdletBinding()]
 param(
+    [ValidateSet("torch", "reservoir")]
+    [string]$Trainer = "torch",
     [string]$Symbols = "",
     [int]$Limit = 0,
     [string]$Lookbacks = "",
     [string]$Units = "",
     [string]$Ridges = "",
+    [string]$Architectures = "",
+    [string]$HiddenSizes = "",
+    [string]$Layers = "",
+    [string]$Dropouts = "",
+    [int]$Epochs = 0,
+    [int]$Patience = 0,
     [string]$Interval = "",
     [string]$EnvFile = ""
 )
@@ -47,9 +55,15 @@ if (-not $Symbols -and $env:LSTM_RETRAIN_SYMBOLS) { $Symbols = $env:LSTM_RETRAIN
 if ($Limit -le 0) {
     $Limit = if ($env:LSTM_RETRAIN_LIMIT) { [int]$env:LSTM_RETRAIN_LIMIT } else { 1000 }
 }
-if (-not $Lookbacks) { $Lookbacks = if ($env:LSTM_RETRAIN_LOOKBACKS) { $env:LSTM_RETRAIN_LOOKBACKS } else { "16,32" } }
+if (-not $Lookbacks) { $Lookbacks = if ($env:LSTM_RETRAIN_LOOKBACKS) { $env:LSTM_RETRAIN_LOOKBACKS } else { "32,64" } }
 if (-not $Units) { $Units = if ($env:LSTM_RETRAIN_UNITS) { $env:LSTM_RETRAIN_UNITS } else { "4,6" } }
 if (-not $Ridges) { $Ridges = if ($env:LSTM_RETRAIN_RIDGES) { $env:LSTM_RETRAIN_RIDGES } else { "0.001" } }
+if (-not $Architectures) { $Architectures = if ($env:LSTM_RETRAIN_ARCHITECTURES) { $env:LSTM_RETRAIN_ARCHITECTURES } else { "lstm,gru" } }
+if (-not $HiddenSizes) { $HiddenSizes = if ($env:LSTM_RETRAIN_HIDDEN_SIZES) { $env:LSTM_RETRAIN_HIDDEN_SIZES } else { "16,32" } }
+if (-not $Layers) { $Layers = if ($env:LSTM_RETRAIN_LAYERS) { $env:LSTM_RETRAIN_LAYERS } else { "1" } }
+if (-not $Dropouts) { $Dropouts = if ($env:LSTM_RETRAIN_DROPOUTS) { $env:LSTM_RETRAIN_DROPOUTS } else { "0.0" } }
+if ($Epochs -le 0) { $Epochs = if ($env:LSTM_RETRAIN_EPOCHS) { [int]$env:LSTM_RETRAIN_EPOCHS } else { 60 } }
+if ($Patience -le 0) { $Patience = if ($env:LSTM_RETRAIN_PATIENCE) { [int]$env:LSTM_RETRAIN_PATIENCE } else { 10 } }
 if (-not $Interval -and $env:LSTM_RETRAIN_INTERVAL) { $Interval = $env:LSTM_RETRAIN_INTERVAL }
 if (-not $EnvFile -and $env:LSTM_RETRAIN_ENV) { $EnvFile = $env:LSTM_RETRAIN_ENV }
 if (-not $EnvFile -and (Test-Path (Join-Path $RepoRoot ".env"))) { $EnvFile = Join-Path $RepoRoot ".env" }
@@ -66,14 +80,29 @@ try {
     }
 
     $python = Resolve-Python
-    $trainerArgs = @(
-        "-u",
-        "tools\train_lstm_forecaster.py",
-        "--limit", $Limit.ToString(),
-        "--lookbacks", $Lookbacks,
-        "--units", $Units,
-        "--ridges", $Ridges
-    )
+    if ($Trainer -eq "torch") {
+        $trainerArgs = @(
+            "-u",
+            "tools\train_torch_recurrent_forecaster.py",
+            "--limit", $Limit.ToString(),
+            "--lookbacks", $Lookbacks,
+            "--architectures", $Architectures,
+            "--hidden-sizes", $HiddenSizes,
+            "--layers", $Layers,
+            "--dropouts", $Dropouts,
+            "--epochs", $Epochs.ToString(),
+            "--patience", $Patience.ToString()
+        )
+    } else {
+        $trainerArgs = @(
+            "-u",
+            "tools\train_lstm_forecaster.py",
+            "--limit", $Limit.ToString(),
+            "--lookbacks", $Lookbacks,
+            "--units", $Units,
+            "--ridges", $Ridges
+        )
+    }
     if ($Symbols) { $trainerArgs += @("--symbols", $Symbols) }
     if ($Interval) { $trainerArgs += @("--interval", $Interval) }
     if ($EnvFile) { $trainerArgs += @("--env", $EnvFile) }

tools/train_torch_recurrent_forecaster.py

@@ -0,0 +1,434 @@
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import sys
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+PROJECT_ROOT = Path(__file__).resolve().parents[1]
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+
+try:
+    import torch
+    from torch import nn
+    from torch.utils.data import DataLoader, TensorDataset
+except ImportError as exc:  # pragma: no cover - exercised on machines without training deps.
+    raise SystemExit(
+        "PyTorch is not installed. Install local training deps with: "
+        "python -m pip install torch --index-url https://download.pytorch.org/whl/cpu"
+    ) from exc
+
+from crypto_spot_bot.bybit import BybitClient
+from crypto_spot_bot.config import load_settings
+from crypto_spot_bot.time_series import _log_returns
+
+
+@dataclass(slots=True)
+class PreparedData:
+    train_x: torch.Tensor
+    train_y: torch.Tensor
+    validation_x: torch.Tensor
+    validation_y: torch.Tensor
+    validation_returns: list[float]
+    mean: float
+    scale: float
+    train_samples: int
+    validation_samples: int
+
+
+class RecurrentReturnModel(nn.Module):
+    def __init__(
+        self,
+        *,
+        architecture: str,
+        hidden_size: int,
+        num_layers: int,
+        dropout: float,
+    ) -> None:
+        super().__init__()
+        recurrent_cls = nn.LSTM if architecture == "lstm" else nn.GRU
+        self.rnn = recurrent_cls(
+            input_size=1,
+            hidden_size=hidden_size,
+            num_layers=num_layers,
+            dropout=dropout if num_layers > 1 else 0.0,
+            batch_first=True,
+        )
+        self.head = nn.Linear(hidden_size, 1)
+
+    def forward(self, values: torch.Tensor) -> torch.Tensor:
+        output, _state = self.rnn(values)
+        return self.head(output[:, -1, :]).squeeze(-1)
+
+
+def main() -> None:
+    args = _parse_args()
+    if args.threads > 0:
+        torch.set_num_threads(args.threads)
+    _seed(args.seed)
+
+    settings = load_settings(args.env)
+    client = BybitClient(settings)
+    symbols = _symbols(args.symbols, settings, client)
+    interval = args.interval or settings.base_interval
+    output = Path(args.output) if args.output else settings.time_series_lstm_model_path
+    device = _device(args.device)
+
+    artifact: dict[str, Any] = {
+        "version": 2,
+        "type": "pytorch_recurrent_forecaster",
+        "created_at": datetime.now(timezone.utc).isoformat(),
+        "trainer": Path(__file__).name,
+        "interval": interval,
+        "limit": args.limit,
+        "validation_window": args.validation_window,
+        "device": str(device),
+        "symbols": {},
+    }
+
+    for symbol in symbols:
+        result = _train_symbol(
+            client=client,
+            symbol=symbol,
+            interval=interval,
+            limit=args.limit,
+            validation_window=args.validation_window,
+            architectures=_strings(args.architectures),
+            lookbacks=_ints(args.lookbacks),
+            hidden_sizes=_ints(args.hidden_sizes),
+            layers_values=_ints(args.layers),
+            dropouts=_floats(args.dropouts),
+            epochs=args.epochs,
+            patience=args.patience,
+            batch_size=args.batch_size,
+            learning_rate=args.learning_rate,
+            weight_decay=args.weight_decay,
+            clip=args.clip,
+            device=device,
+            seed=args.seed,
+        )
+        if result is None:
+            print(f"{symbol}: skipped, not enough candles or train/validation samples")
+            continue
+        artifact["symbols"][symbol] = result
+        print(
+            f"{symbol}: model={result['model']} lookback={result['lookback']} "
+            f"hidden={result['hidden_size']} layers={result['num_layers']} "
+            f"mae={result['validation_mae_percent']:.5f}% "
+            f"baseline={result['baseline_mae_percent']:.5f}% skill={result['skill']:.4f}"
+        )
+
+    output.parent.mkdir(parents=True, exist_ok=True)
+    tmp_output = output.with_name(f"{output.name}.tmp")
+    tmp_output.write_text(json.dumps(artifact, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    tmp_output.replace(output)
+    print(f"saved {output}")
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Train PyTorch LSTM/GRU forecast models on Bybit spot candles.")
+    parser.add_argument("--env", default=None, help="Path to .env file.")
+    parser.add_argument("--symbols", default="", help="Comma-separated symbols. Defaults to configured or popular pairs.")
+    parser.add_argument("--interval", default="", help="Bybit kline interval. Defaults to BASE_INTERVAL.")
+    parser.add_argument("--limit", type=int, default=1000, help="Kline limit per symbol.")
+    parser.add_argument("--validation-window", type=int, default=120, help="Held-out tail returns used for validation.")
+    parser.add_argument("--architectures", default="lstm,gru", help="Comma-separated recurrent types: lstm,gru.")
+    parser.add_argument("--lookbacks", default="32,64", help="Comma-separated sequence lengths.")
+    parser.add_argument("--hidden-sizes", default="16,32", help="Comma-separated hidden sizes.")
+    parser.add_argument("--layers", default="1", help="Comma-separated recurrent layer counts.")
+    parser.add_argument("--dropouts", default="0.0", help="Comma-separated dropout values; only used with layers > 1.")
+    parser.add_argument("--epochs", type=int, default=60, help="Maximum epochs per hyperparameter candidate.")
+    parser.add_argument("--patience", type=int, default=10, help="Early stopping patience in epochs.")
+    parser.add_argument("--batch-size", type=int, default=64, help="Training batch size.")
+    parser.add_argument("--learning-rate", type=float, default=0.001, help="AdamW learning rate.")
+    parser.add_argument("--weight-decay", type=float, default=0.0001, help="AdamW weight decay.")
+    parser.add_argument("--clip", type=float, default=8.0, help="Clamp normalized returns and predictions to this range.")
+    parser.add_argument("--seed", type=int, default=7, help="Random seed.")
+    parser.add_argument("--threads", type=int, default=0, help="Torch CPU threads; 0 keeps torch default.")
+    parser.add_argument("--device", default="auto", help="auto, cpu, cuda, or mps.")
+    parser.add_argument("--output", default="", help="Output JSON path. Defaults to TIME_SERIES_LSTM_MODEL_PATH.")
+    return parser.parse_args()
+
+
+def _symbols(raw: str, settings: Any, client: BybitClient) -> list[str]:
+    if raw.strip():
+        return [item.strip().upper() for item in raw.split(",") if item.strip()]
+    if settings.symbols:
+        return list(settings.symbols)
+    return client.popular_spot_symbols(settings.top_symbols_count)
+
+
+def _train_symbol(
+    *,
+    client: BybitClient,
+    symbol: str,
+    interval: str,
+    limit: int,
+    validation_window: int,
+    architectures: list[str],
+    lookbacks: list[int],
+    hidden_sizes: list[int],
+    layers_values: list[int],
+    dropouts: list[float],
+    epochs: int,
+    patience: int,
+    batch_size: int,
+    learning_rate: float,
+    weight_decay: float,
+    clip: float,
+    device: torch.device,
+    seed: int,
+) -> dict[str, Any] | None:
+    candles = client.klines(symbol, interval, limit)
+    closes = [float(candle.close) for candle in candles if candle.close > 0]
+    returns = _log_returns(closes)
+    if len(returns) < max(100, validation_window + 80):
+        return None
+
+    best: dict[str, Any] | None = None
+    for lookback in lookbacks:
+        prepared = _prepare_data(returns, lookback, validation_window, clip, device)
+        if prepared is None:
+            continue
+        baseline_mae = sum(abs(value) for value in prepared.validation_returns) / len(prepared.validation_returns)
+        for architecture in architectures:
+            if architecture not in {"lstm", "gru"}:
+                continue
+            for hidden_size in hidden_sizes:
+                for num_layers in layers_values:
+                    for dropout in dropouts:
+                        candidate = _fit_candidate(
+                            prepared=prepared,
+                            architecture=architecture,
+                            hidden_size=hidden_size,
+                            num_layers=num_layers,
+                            dropout=dropout,
+                            epochs=epochs,
+                            patience=patience,
+                            batch_size=batch_size,
+                            learning_rate=learning_rate,
+                            weight_decay=weight_decay,
+                            clip=clip,
+                            device=device,
+                            seed=seed,
+                        )
+                        validation_mae = float(candidate["validation_mae"])
+                        skill = (baseline_mae - validation_mae) / baseline_mae if baseline_mae > 0 else 0.0
+                        row = {
+                            **candidate,
+                            "model": f"torch_{architecture}",
+                            "architecture": architecture,
+                            "lookback": lookback,
+                            "hidden_size": hidden_size,
+                            "num_layers": num_layers,
+                            "dropout": dropout,
+                            "mean": prepared.mean,
+                            "scale": prepared.scale,
+                            "clip": clip,
+                            "validation_mae_percent": validation_mae * 100,
+                            "baseline_mae_percent": baseline_mae * 100,
+                            "skill": skill,
+                            "candles": len(candles),
+                            "returns": len(returns),
+                            "train_samples": prepared.train_samples,
+                            "validation_samples": prepared.validation_samples,
+                        }
+                        if best is None or validation_mae < float(best["validation_mae"]):
+                            best = row
+    if best is None:
+        return None
+    best.pop("validation_mae", None)
+    return best
+
+
+def _prepare_data(
+    returns: list[float],
+    lookback: int,
+    validation_window: int,
+    clip: float,
+    device: torch.device,
+) -> PreparedData | None:
+    validation_window = min(max(16, validation_window), max(16, len(returns) // 3))
+    split = len(returns) - validation_window
+    if split <= lookback + 16:
+        return None
+    train_returns = returns[:split]
+    mean = sum(train_returns) / len(train_returns)
+    scale = _return_scale(train_returns)
+    normalized = [_clamp((value - mean) / scale, -clip, clip) for value in returns]
+    train_x: list[list[list[float]]] = []
+    train_y: list[float] = []
+    validation_x: list[list[list[float]]] = []
+    validation_y: list[float] = []
+    validation_returns: list[float] = []
+    for target_index in range(lookback, len(returns)):
+        row = [[value] for value in normalized[target_index - lookback : target_index]]
+        target = normalized[target_index]
+        if target_index < split:
+            train_x.append(row)
+            train_y.append(target)
+        else:
+            validation_x.append(row)
+            validation_y.append(target)
+            validation_returns.append(returns[target_index])
+    if len(train_x) < 24 or len(validation_x) < 8:
+        return None
+    return PreparedData(
+        train_x=torch.tensor(train_x, dtype=torch.float32, device=device),
+        train_y=torch.tensor(train_y, dtype=torch.float32, device=device),
+        validation_x=torch.tensor(validation_x, dtype=torch.float32, device=device),
+        validation_y=torch.tensor(validation_y, dtype=torch.float32, device=device),
+        validation_returns=validation_returns,
+        mean=mean,
+        scale=scale,
+        train_samples=len(train_x),
+        validation_samples=len(validation_x),
+    )
+
+
+def _fit_candidate(
+    *,
+    prepared: PreparedData,
+    architecture: str,
+    hidden_size: int,
+    num_layers: int,
+    dropout: float,
+    epochs: int,
+    patience: int,
+    batch_size: int,
+    learning_rate: float,
+    weight_decay: float,
+    clip: float,
+    device: torch.device,
+    seed: int,
+) -> dict[str, Any]:
+    _seed(seed)
+    model = RecurrentReturnModel(
+        architecture=architecture,
+        hidden_size=hidden_size,
+        num_layers=num_layers,
+        dropout=dropout,
+    ).to(device)
+    optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
+    criterion = nn.SmoothL1Loss(beta=0.5)
+    generator = torch.Generator(device="cpu").manual_seed(seed)
+    loader = DataLoader(
+        TensorDataset(prepared.train_x, prepared.train_y),
+        batch_size=max(1, batch_size),
+        shuffle=True,
+        generator=generator,
+    )
+
+    best_state: dict[str, torch.Tensor] | None = None
+    best_mae = math.inf
+    best_epoch = 0
+    stale_epochs = 0
+    for epoch in range(1, max(1, epochs) + 1):
+        model.train()
+        for batch_x, batch_y in loader:
+            optimizer.zero_grad(set_to_none=True)
+            loss = criterion(model(batch_x), batch_y)
+            loss.backward()
+            nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
+            optimizer.step()
+
+        validation_mae = _validation_mae(model, prepared, clip)
+        if validation_mae + 1e-12 < best_mae:
+            best_mae = validation_mae
+            best_epoch = epoch
+            best_state = {key: value.detach().cpu().clone() for key, value in model.state_dict().items()}
+            stale_epochs = 0
+        else:
+            stale_epochs += 1
+            if stale_epochs >= max(1, patience):
+                break
+
+    if best_state:
+        model.load_state_dict(best_state)
+    return {
+        "validation_mae": best_mae,
+        "best_epoch": best_epoch,
+        "epochs_trained": best_epoch + stale_epochs,
+        "state_dict": _export_recurrent_state(model),
+        "head_weight": _round_list(model.head.weight.detach().cpu().squeeze(0).tolist()),
+        "head_bias": round(float(model.head.bias.detach().cpu().item()), 10),
+    }
+
+
+def _validation_mae(model: nn.Module, prepared: PreparedData, clip: float) -> float:
+    model.eval()
+    with torch.no_grad():
+        normalized_predictions = model(prepared.validation_x).detach().cpu().tolist()
+    errors = []
+    for prediction, actual in zip(normalized_predictions, prepared.validation_returns):
+        raw_prediction = _clamp(float(prediction), -clip, clip) * prepared.scale + prepared.mean
+        errors.append(abs(raw_prediction - actual))
+    return sum(errors) / len(errors) if errors else math.inf
+
+
+def _export_recurrent_state(model: RecurrentReturnModel) -> dict[str, Any]:
+    return {
+        key: _round_nested(value.detach().cpu().tolist())
+        for key, value in model.rnn.state_dict().items()
+    }
+
+
+def _device(raw: str) -> torch.device:
+    value = raw.strip().lower()
+    if value == "auto":
+        if torch.cuda.is_available():
+            return torch.device("cuda")
+        if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available():
+            return torch.device("mps")
+        return torch.device("cpu")
+    return torch.device(value)
+
+
+def _seed(seed: int) -> None:
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+
+def _return_scale(returns: list[float]) -> float:
+    values = sorted(abs(value) for value in returns 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 _clamp(value: float, low: float, high: float) -> float:
+    return max(low, min(high, value))
+
+
+def _round_nested(value: Any) -> Any:
+    if isinstance(value, list):
+        return [_round_nested(item) for item in value]
+    return round(float(value), 10)
+
+
+def _round_list(values: list[float]) -> list[float]:
+    return [round(float(value), 10) for value in values]
+
+
+def _ints(raw: str) -> list[int]:
+    return [int(item.strip()) for item in raw.split(",") if item.strip()]
+
+
+def _floats(raw: str) -> list[float]:
+    return [float(item.strip()) for item in raw.split(",") if item.strip()]
+
+
+def _strings(raw: str) -> list[str]:
+    return [item.strip().lower() for item in raw.split(",") if item.strip()]
+
+
+if __name__ == "__main__":
+    main()