diff --git a/alpha/AutoRegression/ar_model.pth b/alpha/AutoRegression/ar_model.pth
new file mode 100644
index 0000000..8efc6fa
Binary files /dev/null and b/alpha/AutoRegression/ar_model.pth differ
diff --git a/alpha/AutoRegression/autoregression.py b/alpha/AutoRegression/autoregression.py
new file mode 100644
index 0000000..2b5b5e9
--- /dev/null
+++ b/alpha/AutoRegression/autoregression.py
@@ -0,0 +1,145 @@
+import torch
+from torch import nn
+from torch.utils.data import Dataset, DataLoader
+import numpy as np
+from pathlib import Path
+
+class AR(nn.Module):
+    def __init__(self, lag):
+        super().__init__()
+        self.lag = lag
+        self.linear = nn.Linear(lag, 1, bias=True)
+        
+    def forward(self, x):
+        return self.linear(x)
+    
+def train_one_epoch(loader: DataLoader, model: nn.Module, loss_fn, optimizer):
+    model.train()
+    size = len(loader.dataset)
+    running_loss = 0.0
+    for _, (xb, yb) in enumerate(loader):
+        xb = xb.float()
+        yb = yb.float()
+        optimizer.zero_grad()
+        pred = model(xb)
+        loss = loss_fn(pred, yb)
+        loss.backward()
+        optimizer.step()
+        running_loss += loss.item() * xb.size(0)
+    epoch_loss = running_loss / size
+    print(f"Average training loss: {epoch_loss: .10e}")
+    return epoch_loss
+
+def train(train_dataset: Dataset, epoches: int, ar_model: nn.Module):
+    loss_fn = nn.L1Loss()
+    optimizer = torch.optim.Adam(ar_model.parameters(), lr=1e-3)
+    loader = DataLoader(dataset=train_dataset, batch_size=64, shuffle=False)
+    for _ in range(epoches):
+        train_one_epoch(loader, ar_model, loss_fn, optimizer)
+
+def build_train(data, T):
+    X = []
+    y = []
+    n = len(data)
+    for i in range(n - T):
+        window = data[i : i + T]
+        target = data[i + T]
+        X.append(window)
+        y.append(target)
+        
+    return np.array(X), np.array(y)
+
+class ArDataset(Dataset):
+    def __init__(self, X, y):
+        X = torch.tensor(X, dtype=torch.float32)
+        y = torch.tensor(y, dtype=torch.float32)
+        self.X = X.float()
+        self.y = y.float()
+        self.y = self.y.unsqueeze(1)
+        
+    def __len__(self):
+        return self.X.shape[0]
+    
+    def __getitem__(self, idx):
+        return self.X[idx], self.y[idx]
+    
+def infer_mu(model, X):
+    model.eval()
+    with torch.no_grad():
+        X_tensor = torch.tensor(X, dtype=torch.float32)
+        mu = model(X_tensor).squeeze(1).cpu().numpy()
+    return mu
+
+def load_ar_model(model_path="./ar_model.pth", lag=90):
+    model = AR(lag=lag)
+    model.load_state_dict(torch.load(model_path, weights_only=True))
+    model.eval()
+    return model
+
+base_dir = Path(__file__).resolve().parent
+model_path = base_dir / "ar_model.pth"
+
+
+class _LazyARModel:
+    """
+    Lazily loads and caches the AR model on first use.
+
+    This avoids loading the model at import time, so that import errors
+    due to missing/corrupt weights or missing dependencies do not
+    prevent the rest of the package from being used.
+    """
+
+    def __init__(self, path):
+        self._path = path
+        self._model = None
+
+    def _load(self):
+        if self._model is None:
+            # Delegate to the existing loader; any exceptions will be
+            # raised at first use rather than at import time.
+            self._model = load_ar_model(self._path)
+        return self._model
+
+    def __getattr__(self, name):
+        # Proxy attribute access to the underlying model instance.
+        return getattr(self._load(), name)
+    
+    def __call__(self, *args, **kwargs):
+        # Forward calls to the underlying model
+        return self._load()(*args, **kwargs)
+
+
+trained_ar_model = _LazyARModel(model_path)
+# if __name__ == "__main__":
+#     print("1. Define model")
+#     ar_model = AR(lag=90)
+    
+#     print("2. Load data")
+#     data = load_data()
+#     log_ret = data["log_return"].to_numpy()
+#     log_ret = log_ret[np.isfinite(log_ret)]
+#     X, y = build_train(log_ret, 90)
+#     print(X[0], y[0])
+    
+#     print("3. Prepare training")
+#     train_idx = int(len(X) * 0.85)
+#     X_train = X[:train_idx]
+#     y_train = y[:train_idx]
+#     dataset = ArDataset(X_train, y_train)
+    
+#     print("4. Train")
+#     train(dataset, 10, ar_model)
+    
+#     print("5. Test inference")
+#     ar_model.eval()
+#     X_test = X[train_idx:]
+#     y_test = y[train_idx:]
+
+#     mu_test = infer_mu(ar_model, X_test)
+#     residuals = y_test - mu_test
+
+#     mae = np.mean(np.abs(residuals))
+#     print("Test MAE:", mae)
+    
+#     print("6. Save model")
+#     torch.save(ar_model.state_dict(), "./ar_model.pth")
\ No newline at end of file
diff --git a/alpha/AutoRegression/data_processing.py b/alpha/AutoRegression/data_processing.py
new file mode 100644
index 0000000..51144f9
--- /dev/null
+++ b/alpha/AutoRegression/data_processing.py
@@ -0,0 +1,16 @@
+import numpy as np
+import pandas as pd
+
+def load_data(data_path: str = "./binance/BTC_USDT-5m.feather",
+              start_date: str = "2021-01-01"):
+    df = pd.read_feather(data_path)
+    df = df[df["date"] > start_date]
+    df["log_return"] = np.log(df['close'] / df['close'].shift(1))
+    print(df.head())
+    return df
+
+def get_log_return_series():
+    df = load_data()
+    log_ret = df["log_return"].to_numpy()
+    log_ret = log_ret[np.isfinite(log_ret)]
+    return log_ret
\ No newline at end of file
diff --git a/alpha/AutoregressionAlpha.py b/alpha/AutoregressionAlpha.py
new file mode 100644
index 0000000..30fd58e
--- /dev/null
+++ b/alpha/AutoregressionAlpha.py
@@ -0,0 +1,32 @@
+import numpy as np
+import torch
+from pandas import DataFrame
+from alpha.interface import IAlpha
+from alpha.AutoRegression.autoregression import trained_ar_model, build_train
+
+AR_MODEL_PATH = "./ar_model.pth"
+AR_LAG = 90
+
+class AutoregressionAlpha(IAlpha):
+    def process(self) -> DataFrame:
+        df = self.dataframe.copy()
+        
+        df["log_return"] = np.log(df["close"] / df["close"].shift(1))
+        log_ret = df["log_return"].to_numpy()
+        log_ret = log_ret[np.isfinite(log_ret)]
+        
+        if len(log_ret) <= AR_LAG:
+            df["ar_pred"] = np.nan
+            return df
+        X, _ = build_train(log_ret, AR_LAG)
+        # Load AR model
+        ar_model = trained_ar_model
+        ar_model.eval()
+        with torch.no_grad():
+            X_tensor = torch.tensor(X, dtype=torch.float32)
+            preds = ar_model(X_tensor).squeeze(1).cpu().numpy()
+        # Align predictions with dataframe index
+        ar_pred = np.full(df.shape[0], np.nan)
+        ar_pred[AR_LAG+1:] = preds
+        df["ar_pred"] = ar_pred
+        return df
\ No newline at end of file
diff --git a/alpha/interface.py b/alpha/interface.py
index d97a052..8652fe3 100644
--- a/alpha/interface.py
+++ b/alpha/interface.py
@@ -12,4 +12,21 @@ def process(self) -> DataFrame:
         This is to decouple the pipulation of indicator from IStrategy
         """
         pass
-        
\ No newline at end of file
+
+fwd_ret_timeframe = [1, 5, 10, 20, 90]
+class AlphaEvaluator:
+    def __init__(self, dataframe: DataFrame, alpha: type[IAlpha], metadata: dict = None):
+        self.df = dataframe
+        self.alpha = alpha(dataframe, metadata if metadata is not None else {})
+        
+    def evaluate_information_coefficient(self, alpha_names):
+        df_processed = self.alpha.process()
+        out = {}
+        for a in alpha_names:
+            for t in fwd_ret_timeframe:
+                fwd_ret = df_processed['close'].pct_change().shift(-t)
+                temp_df = DataFrame({'alpha': df_processed[a], 'fwd_ret': fwd_ret})
+                temp_df = temp_df.dropna(subset=['alpha', 'fwd_ret'])
+                ic = temp_df['alpha'].corr(temp_df['fwd_ret'], method='spearman')
+                out[(a, t)] = ic
+        return out
\ No newline at end of file
diff --git a/freqtrade b/freqtrade
index 8c7385d..d57e6c5 160000
--- a/freqtrade
+++ b/freqtrade
@@ -1 +1 @@
-Subproject commit 8c7385dc98ecab1359edd878afcff91fffdec683
+Subproject commit d57e6c5e5bca53cfc63ee10063710bf7e46d8df0
diff --git a/tests/test_alpha.py b/tests/test_alpha.py
index 12117a8..efdcdc6 100644
--- a/tests/test_alpha.py
+++ b/tests/test_alpha.py
@@ -6,11 +6,12 @@
 
 talib = pytest.importorskip("talib", reason="TA-Lib C library not installed")
 
-from alpha.interface import IAlpha
+from alpha.interface import IAlpha, AlphaEvaluator
 from alpha.SimpleEmaFactors import EmaAlpha
 from alpha.RsiAlpha import RsiAlpha
 from alpha.MacdAlpha import MacdAlpha
 from alpha.BollingerAlpha import BollingerAlpha
+from alpha.AutoregressionAlpha import AutoregressionAlpha
 from alpha.EventLstmAlpha import EventLstmAlpha # type: ignore
 
 def _make_ohlcv(n=100):
@@ -160,6 +161,62 @@ def test_pctb_near_0_5_for_mean(self):
         # Mean %B across a random walk should be roughly centered around 0.5
         assert 0.2 < valid.mean() < 0.8
 
+class TestAutoregressionAlpha:
+    def test_process_adds_ar_pred_column(self):
+        df = _make_ohlcv(200)
+        df = df[["date", "close"]].copy()
+        result = AutoregressionAlpha(df).process()
+        assert "ar_pred" in result.columns, "Missing ar_pred column"
+
+    def test_ar_pred_nan_for_short_series(self):
+        df = _make_ohlcv(50)
+        df = df[["date", "close"]].copy()
+        result = AutoregressionAlpha(df).process()
+        assert result["ar_pred"].isna().all(), "ar_pred should be NaN for short series"
+
+    def test_ar_pred_has_valid_values(self):
+        df = _make_ohlcv(200)
+        df = df[["date", "close"]].copy()
+        result = AutoregressionAlpha(df).process()
+        ar_pred = result["ar_pred"].iloc[91:]
+        assert ar_pred.notna().any(), "ar_pred should have valid predictions after lag"
+
+class AlphaStub(IAlpha):
+    def process(self):
+        df = self.dataframe.copy()
+        df["alpha"] = df["close"].pct_change()
+        return df
+
+class TestAlphaInformationEvaluation:
+    def test_evaluate_information_coefficient_returns_dict(self):
+        df = _make_ohlcv(120)
+        evaluator = AlphaEvaluator(df, AlphaStub)
+        result = evaluator.evaluate_information_coefficient(["alpha"])
+        assert isinstance(result, dict)
+        expected_keys = [("alpha", t) for t in [1, 5, 10, 20, 90]]
+        for key in expected_keys:
+            assert key in result
+            assert isinstance(result[key], (float, type(None)))
+
+    def test_ic_values_are_finite_or_nan(self):
+        df = _make_ohlcv(120)
+        evaluator = AlphaEvaluator(df, AlphaStub)
+        result = evaluator.evaluate_information_coefficient(["alpha"])
+        for ic in result.values():
+            assert (ic is None) or (isinstance(ic, float))
+
+    def test_ic_on_constant_alpha_is_nan(self):
+        class ConstAlpha(IAlpha):
+            def process(self):
+                df = self.dataframe.copy()
+                df["alpha"] = 1.0
+                return df
+        df = _make_ohlcv(120)
+        evaluator = AlphaEvaluator(df, ConstAlpha)
+        result = evaluator.evaluate_information_coefficient(["alpha"])
+        for ic in result.values():
+            assert ic != ic
+
 class TestEventLstmAlpha:
     @pytest.mark.slow
     def test_event_lstm_alpha_with_fake_data(self):
@@ -196,4 +253,4 @@ def test_event_lstm_alpha_with_fake_data(self):
         assert valid_preds.notna().all(), "Predictions should be filled"
         
         # Predictions should be within [0,1] because close prices are normalized
-        assert (valid_preds >= 0).all() and (valid_preds <= 1).all()
\ No newline at end of file
+        assert (valid_preds >= 0).all() and (valid_preds <= 1).all()