baonathor
/

xtrade-v1

@@ -1,333 +1,333 @@
-import torch
-import torch.nn as nn
-import numpy as np
-import pandas as pd
-import math
-import os
-import joblib
-import time
-from typing import List, Dict, Optional
-from transformers import AutoModel, AutoTokenizer, AutoConfig
-# ==========================================
-# 1. CONFIGURATION
-# ==========================================
-class PublicConfig:
-    # Model Architecture
-    max_length = 256
-    num_labels_3m = 3
-    num_labels_30m = 3
-    # Feature settings
-    feature_cols = [
-        "feat_ret_1m", "feat_ret_5m", "feat_ret_15m",
-        "feat_volatility_60m", "feat_num_trades_60m", "feat_volume_60m",
-        "feat_tweet_freq_24h", "feat_time_since_prev_tweet",
-        "feat_btc_ret_60m", "feat_btc_ret_24h",
-        "feat_fear_greed_index", "feat_btc_dominance", "feat_altseason_index"
-    ]
-    # Inference Settings
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-    # Paths (Relative to this script or defined by user)
-    checkpoint_dir = "checkpoints_market_event_multitask"
-    model_filename = "best_model_extreme.pt"
-    scaler_filename = "scaler.pkl"
-cfg = PublicConfig()
-# ==========================================
-# 2. MODEL ARCHITECTURE
-# ==========================================
-class MarketConditionedEventMultiTask(nn.Module):
-    """
-    The Extreme Signal Model Architecture.
-    Combines BERT (Text) + MLP (Market Data) + Attention Mechanism.
-    """
-    def __init__(self, num_features: int,
-                 num_labels_3m: int, num_labels_30m: int,
-                 bert_config, device: str = "cpu"):
-        super().__init__()
-        # Load BERT structure from config (Offline mode)
-        self.bert = AutoModel.from_config(bert_config)
-        hidden_size = self.bert.config.hidden_size
-        # MLP to encode numeric market features
-        self.market_mlp = nn.Sequential(
-            nn.Linear(num_features, hidden_size),
-            nn.ReLU(),
-            nn.LayerNorm(hidden_size),
-        )
-        # Linear projections for market-conditioned attention
-        self.query_proj = nn.Linear(hidden_size, hidden_size)
-        self.key_proj = nn.Linear(hidden_size, hidden_size)
-        combined_size = hidden_size * 3  # [CLS] + context + market_emb
-        # Classification head for 3m horizon
-        self.classifier_3m = nn.Sequential(
-            nn.Linear(combined_size, hidden_size),
-            nn.ReLU(),
-            nn.Dropout(0.2),
-            nn.Linear(hidden_size, num_labels_3m),
-        )
-        # Classification head for 30m horizon
-        self.classifier_30m = nn.Sequential(
-            nn.Linear(combined_size, hidden_size),
-            nn.ReLU(),
-            nn.Dropout(0.2),
-            nn.Linear(hidden_size, num_labels_30m),
-        )
-    def forward(self, input_ids, attention_mask, market_features):
-        # Encode tweet with BERT
-        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
-        last_hidden_state = outputs.last_hidden_state
-        pooled_output = outputs.pooler_output
-        # Encode market state
-        market_emb = self.market_mlp(market_features)
-        # Market-conditioned attention
-        Q = self.query_proj(market_emb).unsqueeze(1)
-        K = self.key_proj(last_hidden_state)
-        scores = torch.matmul(Q, K.transpose(1, 2)) / math.sqrt(K.size(-1))
-        extended_mask = attention_mask.unsqueeze(1)
-        scores = scores.masked_fill(extended_mask == 0, float("-inf"))
-        attn_weights = torch.softmax(scores, dim=-1)
-        context = torch.matmul(attn_weights, last_hidden_state).squeeze(1)
-        # Combine and Classify
-        combined = torch.cat([pooled_output, context, market_emb], dim=-1)
-        logits_3m = self.classifier_3m(combined)
-        logits_30m = self.classifier_30m(combined)
-        return logits_3m, logits_30m
-# ==========================================
-# 3. INFERENCE CLASS
-# ==========================================
-class ExtremeModelPredictor:
-    def __init__(self, model_dir: str):
-        self.device = cfg.device
-        self.model_dir = model_dir
-        print(f"Loading Extreme Model from {model_dir}...")
-        print(f"Using device: {self.device.upper()}")
-        if self.device == 'cpu':
-            print("Note: Running on CPU. If you have an NVIDIA GPU, please install PyTorch with CUDA support.")
-        # 1. Load Tokenizer
-        try:
-            self.tokenizer = AutoTokenizer.from_pretrained(model_dir)
-        except Exception as e:
-            raise FileNotFoundError(f"Could not load tokenizer from {model_dir}. Please ensure tokenizer files exist. Error: {e}")
-        # 2. Load Scaler
-        scaler_path = os.path.join(model_dir, cfg.scaler_filename)
-        if not os.path.exists(scaler_path):
-            # Fallback for development environment
-            scaler_path = cfg.scaler_filename
-        if os.path.exists(scaler_path):
-            self.scaler = joblib.load(scaler_path)
-        else:
-            raise FileNotFoundError(f"Scaler not found at {scaler_path}. Please include scaler.pkl.")
-        # 3. Load Model Config & Weights
-        config_path = os.path.join(model_dir, "config.json")
-        if not os.path.exists(config_path):
-            raise FileNotFoundError(f"Config not found at {config_path}. Please ensure config.json exists.")
-        bert_config = AutoConfig.from_pretrained(model_dir)
-        self.model = MarketConditionedEventMultiTask(
-            num_features=len(cfg.feature_cols),
-            num_labels_3m=cfg.num_labels_3m,
-            num_labels_30m=cfg.num_labels_30m,
-            device=self.device,
-            bert_config=bert_config
-        )
-        # Load State Dict
-        weight_path = os.path.join(model_dir, cfg.model_filename)
-        if not os.path.exists(weight_path):
-             # Fallback name
-             weight_path = os.path.join(model_dir, "best_model.pt")
-        state_dict = torch.load(weight_path, map_location="cpu")
-        self.model.load_state_dict(state_dict)
-        self.model.to(self.device)
-        self.model.eval()
-        print("Model loaded successfully.")
-    def preprocess_features(self, raw_feats: Dict[str, float]) -> np.ndarray:
-        # Ensure correct order and fill missing with 0
-        vals = [raw_feats.get(col, 0.0) for col in cfg.feature_cols]
-        # Create DataFrame with feature names to avoid sklearn warning
-        df = pd.DataFrame([vals], columns=cfg.feature_cols)
-        df = df.fillna(0)
-        return self.scaler.transform(df)
-    def predict(self, project_name: str, symbol: str, tweet_text: str, market_features: Dict[str, float]):
-        """
-        Returns probabilities for 3m and 30m horizons.
-        Classes: 0 (Down), 1 (Neutral), 2 (Up)
-        """
-        full_text = f"{project_name} ({symbol}): {tweet_text}"
-        start_time = time.perf_counter()
-        # Tokenize
-        encoded = self.tokenizer(
-            full_text,
-            padding="max_length",
-            truncation=True,
-            max_length=cfg.max_length,
-            return_tensors="pt"
-        )
-        input_ids = encoded["input_ids"].to(self.device)
-        attention_mask = encoded["attention_mask"].to(self.device)
-        # Features
-        feats_scaled = self.preprocess_features(market_features)
-        feats_tensor = torch.tensor(feats_scaled, dtype=torch.float32).to(self.device)
-        with torch.no_grad():
-            logits_3m, logits_30m = self.model(input_ids, attention_mask, feats_tensor)
-            probs_3m = torch.softmax(logits_3m, dim=-1).cpu().numpy()[0]
-            probs_30m = torch.softmax(logits_30m, dim=-1).cpu().numpy()[0]
-        inference_time = time.perf_counter() - start_time
-        return {
-            "3m_probs": {"Down": probs_3m[0], "Neutral": probs_3m[1], "Up": probs_3m[2]},
-            "30m_probs": {"Down": probs_30m[0], "Neutral": probs_30m[1], "Up": probs_30m[2]},
-            "extreme_signal": self._check_extreme(probs_3m),
-            "inference_time": inference_time
-        }
-    def _check_extreme(self, probs):
-        # Logic: If Down or Up > 0.7
-        threshold = 0.7
-        if probs[0] > threshold:
-            return "EXTREME DOWN"
-        elif probs[2] > threshold:
-            return "EXTREME UP"
-        return "NORMAL"
-# ==========================================
-# 4. EXAMPLE USAGE
-# ==========================================
-if __name__ == "__main__":
-    # Example of how to use this script
-    # Path to the folder containing: config.json, tokenizer files, scaler.pkl, best_model_extreme.pt
-    checkpoint_folder = "checkpoints_market_event_multitask"
-    # Check if folder exists before running
-    if os.path.exists(checkpoint_folder):
-        predictor = ExtremeModelPredictor(checkpoint_folder)
-        # Define a helper to run scenarios
-        def run_scenario(name, tweet, features):
-            print(f"\n--- Scenario: {name} ---")
-            print(f"Tweet: {tweet}")
-            # Print key features for context
-            print(f"Key Feats: 1m={features.get('feat_ret_1m')}, Vol={features.get('feat_volume_60m')}, F&G={features.get('feat_fear_greed_index')}")
-            res = predictor.predict("Test", "TST", tweet, features)
-            print(f"Signal: {res['extreme_signal']}")
-            print(f"3m Probs: Down={res['3m_probs']['Down']:.3f}, Neutral={res['3m_probs']['Neutral']:.3f}, Up={res['3m_probs']['Up']:.3f}")
-            print(f"Time: {res['inference_time']*1000:.2f} ms")
-            return res
-        # Warm-up
-        print("\n--- Warm-up Run ---")
-        predictor.predict("Warmup", "WP", "Warmup", {
-            "feat_ret_1m": 0.0, "feat_ret_5m": 0.0, "feat_ret_15m": 0.0,
-            "feat_volatility_60m": 0.0, "feat_num_trades_60m": 0, "feat_volume_60m": 0,
-            "feat_tweet_freq_24h": 0, "feat_time_since_prev_tweet": 0,
-            "feat_btc_ret_60m": 0.0, "feat_btc_ret_24h": 0.0,
-            "feat_fear_greed_index": 50, "feat_btc_dominance": 50, "feat_altseason_index": 0
-        })
-        # 1. FOMO / Strong Uptrend
-        run_scenario(
-            "FOMO / Strong Uptrend",
-            "BREAKING: Major exchange listing confirmed! 🚀 #ToTheMoon",
-            {
-                "feat_ret_1m": 0.02, "feat_ret_5m": 0.05, "feat_ret_15m": 0.08,
-                "feat_volatility_60m": 0.05, "feat_num_trades_60m": 1000, "feat_volume_60m": 2000000,
-                "feat_tweet_freq_24h": 100, "feat_time_since_prev_tweet": 5,
-                "feat_btc_ret_60m": 0.01, "feat_btc_ret_24h": 0.05,
-                "feat_fear_greed_index": 80, "feat_btc_dominance": 45, "feat_altseason_index": 80
-            }
-        )
-        # 2. Panic Dump / Crash (The "Mean Reversion" Test)
-        run_scenario(
-            "Panic Dump / Crash",
-            "URGENT: Security breach detected. Do not interact with contracts.",
-            {
-                "feat_ret_1m": -0.05, "feat_ret_5m": -0.08, "feat_ret_15m": -0.10,
-                "feat_volatility_60m": 0.15, "feat_num_trades_60m": 2000, "feat_volume_60m": 5000000,
-                "feat_tweet_freq_24h": 200, "feat_time_since_prev_tweet": 1,
-                "feat_btc_ret_60m": -0.03, "feat_btc_ret_24h": -0.08,
-                "feat_fear_greed_index": 10, "feat_btc_dominance": 60, "feat_altseason_index": 5
-            }
-        )
-        # 3. Slow Bleed / Bear Market
-        run_scenario(
-            "Slow Bleed / Bear Market",
-            "Weekly development update. Progress is slow but steady.",
-            {
-                "feat_ret_1m": -0.001, "feat_ret_5m": -0.002, "feat_ret_15m": -0.005,
-                "feat_volatility_60m": 0.01, "feat_num_trades_60m": 50, "feat_volume_60m": 100000,
-                "feat_tweet_freq_24h": 10, "feat_time_since_prev_tweet": 60,
-                "feat_btc_ret_60m": -0.001, "feat_btc_ret_24h": -0.01,
-                "feat_fear_greed_index": 30, "feat_btc_dominance": 55, "feat_altseason_index": 10
-            }
-        )
-        # 4. Sideways / Stable
-        run_scenario(
-            "Sideways / Stable",
-            "Just a normal day building. #Crypto",
-            {
-                "feat_ret_1m": 0.000, "feat_ret_5m": 0.001, "feat_ret_15m": -0.001,
-                "feat_volatility_60m": 0.005, "feat_num_trades_60m": 20, "feat_volume_60m": 50000,
-                "feat_tweet_freq_24h": 5, "feat_time_since_prev_tweet": 120,
-                "feat_btc_ret_60m": 0.000, "feat_btc_ret_24h": 0.002,
-                "feat_fear_greed_index": 50, "feat_btc_dominance": 50, "feat_altseason_index": 20
-            }
-        )
-        # 5. Divergence: Good News + Bad Price (Opportunity?)
-        run_scenario(
-            "Divergence: Good News + Bad Price",
-            "Partnership with Google Cloud announced!",
-            {
-                "feat_ret_1m": -0.02, "feat_ret_5m": -0.03, "feat_ret_15m": -0.03,
-                "feat_volatility_60m": 0.04, "feat_num_trades_60m": 150, "feat_volume_60m": 300000,
-                "feat_tweet_freq_24h": 20, "feat_time_since_prev_tweet": 10,
-                "feat_btc_ret_60m": -0.01, "feat_btc_ret_24h": -0.02,
-                "feat_fear_greed_index": 40, "feat_btc_dominance": 52, "feat_altseason_index": 15
-            }
-        )
-    else:
-        print(f"Error: Checkpoint folder '{checkpoint_folder}' not found.")
-        print("Please place this script next to your model folder.")

+import torch
+import torch.nn as nn
+import numpy as np
+import pandas as pd
+import math
+import os
+import joblib
+import time
+from typing import List, Dict, Optional
+from transformers import AutoModel, AutoTokenizer, AutoConfig
+# ==========================================
+# 1. CONFIGURATION
+# ==========================================
+class PublicConfig:
+    # Model Architecture
+    max_length = 256
+    num_labels_3m = 3
+    num_labels_30m = 3
+    # Feature settings
+    feature_cols = [
+        "feat_ret_1m", "feat_ret_5m", "feat_ret_15m",
+        "feat_volatility_60m", "feat_num_trades_60m", "feat_volume_60m",
+        "feat_tweet_freq_24h", "feat_time_since_prev_tweet",
+        "feat_btc_ret_60m", "feat_btc_ret_24h",
+        "feat_fear_greed_index", "feat_btc_dominance", "feat_altseason_index"
+    ]
+    # Inference Settings
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    # Paths (Relative to this script or defined by user)
+    checkpoint_dir = "checkpoints_market_event_multitask"
+    model_filename = "best_model_extreme.pt"
+    scaler_filename = "scaler.pkl"
+cfg = PublicConfig()
+# ==========================================
+# 2. MODEL ARCHITECTURE
+# ==========================================
+class MarketConditionedEventMultiTask(nn.Module):
+    """
+    The Extreme Signal Model Architecture.
+    Combines BERT (Text) + MLP (Market Data) + Attention Mechanism.
+    """
+    def __init__(self, num_features: int,
+                 num_labels_3m: int, num_labels_30m: int,
+                 bert_config, device: str = "cpu"):
+        super().__init__()
+        # Load BERT structure from config (Offline mode)
+        self.bert = AutoModel.from_config(bert_config)
+        hidden_size = self.bert.config.hidden_size
+        # MLP to encode numeric market features
+        self.market_mlp = nn.Sequential(
+            nn.Linear(num_features, hidden_size),
+            nn.ReLU(),
+            nn.LayerNorm(hidden_size),
+        )
+        # Linear projections for market-conditioned attention
+        self.query_proj = nn.Linear(hidden_size, hidden_size)
+        self.key_proj = nn.Linear(hidden_size, hidden_size)
+        combined_size = hidden_size * 3  # [CLS] + context + market_emb
+        # Classification head for 3m horizon
+        self.classifier_3m = nn.Sequential(
+            nn.Linear(combined_size, hidden_size),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(hidden_size, num_labels_3m),
+        )
+        # Classification head for 30m horizon
+        self.classifier_30m = nn.Sequential(
+            nn.Linear(combined_size, hidden_size),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(hidden_size, num_labels_30m),
+        )
+    def forward(self, input_ids, attention_mask, market_features):
+        # Encode tweet with BERT
+        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
+        last_hidden_state = outputs.last_hidden_state
+        pooled_output = outputs.pooler_output
+        # Encode market state
+        market_emb = self.market_mlp(market_features)
+        # Market-conditioned attention
+        Q = self.query_proj(market_emb).unsqueeze(1)
+        K = self.key_proj(last_hidden_state)
+        scores = torch.matmul(Q, K.transpose(1, 2)) / math.sqrt(K.size(-1))
+        extended_mask = attention_mask.unsqueeze(1)
+        scores = scores.masked_fill(extended_mask == 0, float("-inf"))
+        attn_weights = torch.softmax(scores, dim=-1)
+        context = torch.matmul(attn_weights, last_hidden_state).squeeze(1)
+        # Combine and Classify
+        combined = torch.cat([pooled_output, context, market_emb], dim=-1)
+        logits_3m = self.classifier_3m(combined)
+        logits_30m = self.classifier_30m(combined)
+        return logits_3m, logits_30m
+# ==========================================
+# 3. INFERENCE CLASS
+# ==========================================
+class ExtremeModelPredictor:
+    def __init__(self, model_dir: str):
+        self.device = cfg.device
+        self.model_dir = model_dir
+        print(f"Loading Extreme Model from {model_dir}...")
+        print(f"Using device: {self.device.upper()}")
+        if self.device == 'cpu':
+            print("Note: Running on CPU. If you have an NVIDIA GPU, please install PyTorch with CUDA support.")
+        # 1. Load Tokenizer
+        try:
+            self.tokenizer = AutoTokenizer.from_pretrained(model_dir)
+        except Exception as e:
+            raise FileNotFoundError(f"Could not load tokenizer from {model_dir}. Please ensure tokenizer files exist. Error: {e}")
+        # 2. Load Scaler
+        scaler_path = os.path.join(model_dir, cfg.scaler_filename)
+        if not os.path.exists(scaler_path):
+            # Fallback for development environment
+            scaler_path = cfg.scaler_filename
+        if os.path.exists(scaler_path):
+            self.scaler = joblib.load(scaler_path)
+        else:
+            raise FileNotFoundError(f"Scaler not found at {scaler_path}. Please include scaler.pkl.")
+        # 3. Load Model Config & Weights
+        config_path = os.path.join(model_dir, "config.json")
+        if not os.path.exists(config_path):
+            raise FileNotFoundError(f"Config not found at {config_path}. Please ensure config.json exists.")
+        bert_config = AutoConfig.from_pretrained(model_dir)
+        self.model = MarketConditionedEventMultiTask(
+            num_features=len(cfg.feature_cols),
+            num_labels_3m=cfg.num_labels_3m,
+            num_labels_30m=cfg.num_labels_30m,
+            device=self.device,
+            bert_config=bert_config
+        )
+        # Load State Dict
+        weight_path = os.path.join(model_dir, cfg.model_filename)
+        if not os.path.exists(weight_path):
+             # Fallback name
+             weight_path = os.path.join(model_dir, "best_model.pt")
+        state_dict = torch.load(weight_path, map_location="cpu",weights_only=False)
+        self.model.load_state_dict(state_dict)
+        self.model.to(self.device)
+        self.model.eval()
+        print("Model loaded successfully.")
+    def preprocess_features(self, raw_feats: Dict[str, float]) -> np.ndarray:
+        # Ensure correct order and fill missing with 0
+        vals = [raw_feats.get(col, 0.0) for col in cfg.feature_cols]
+        # Create DataFrame with feature names to avoid sklearn warning
+        df = pd.DataFrame([vals], columns=cfg.feature_cols)
+        df = df.fillna(0)
+        return self.scaler.transform(df)
+    def predict(self, project_name: str, symbol: str, tweet_text: str, market_features: Dict[str, float]):
+        """
+        Returns probabilities for 3m and 30m horizons.
+        Classes: 0 (Down), 1 (Neutral), 2 (Up)
+        """
+        full_text = f"{project_name} ({symbol}): {tweet_text}"
+        start_time = time.perf_counter()
+        # Tokenize
+        encoded = self.tokenizer(
+            full_text,
+            padding="max_length",
+            truncation=True,
+            max_length=cfg.max_length,
+            return_tensors="pt"
+        )
+        input_ids = encoded["input_ids"].to(self.device)
+        attention_mask = encoded["attention_mask"].to(self.device)
+        # Features
+        feats_scaled = self.preprocess_features(market_features)
+        feats_tensor = torch.tensor(feats_scaled, dtype=torch.float32).to(self.device)
+        with torch.no_grad():
+            logits_3m, logits_30m = self.model(input_ids, attention_mask, feats_tensor)
+            probs_3m = torch.softmax(logits_3m, dim=-1).cpu().numpy()[0]
+            probs_30m = torch.softmax(logits_30m, dim=-1).cpu().numpy()[0]
+        inference_time = time.perf_counter() - start_time
+        return {
+            "3m_probs": {"Down": probs_3m[0], "Neutral": probs_3m[1], "Up": probs_3m[2]},
+            "30m_probs": {"Down": probs_30m[0], "Neutral": probs_30m[1], "Up": probs_30m[2]},
+            "extreme_signal": self._check_extreme(probs_3m),
+            "inference_time": inference_time
+        }
+    def _check_extreme(self, probs):
+        # Logic: If Down or Up > 0.7
+        threshold = 0.7
+        if probs[0] > threshold:
+            return "EXTREME DOWN"
+        elif probs[2] > threshold:
+            return "EXTREME UP"
+        return "NORMAL"
+# ==========================================
+# 4. EXAMPLE USAGE
+# ==========================================
+if __name__ == "__main__":
+    # Example of how to use this script
+    # Path to the folder containing: config.json, tokenizer files, scaler.pkl, best_model_extreme.pt
+    checkpoint_folder = "checkpoints_market_event_multitask"
+    # Check if folder exists before running
+    if os.path.exists(checkpoint_folder):
+        predictor = ExtremeModelPredictor(checkpoint_folder)
+        # Define a helper to run scenarios
+        def run_scenario(name, tweet, features):
+            print(f"\n--- Scenario: {name} ---")
+            print(f"Tweet: {tweet}")
+            # Print key features for context
+            print(f"Key Feats: 1m={features.get('feat_ret_1m')}, Vol={features.get('feat_volume_60m')}, F&G={features.get('feat_fear_greed_index')}")
+            res = predictor.predict("Test", "TST", tweet, features)
+            print(f"Signal: {res['extreme_signal']}")
+            print(f"3m Probs: Down={res['3m_probs']['Down']:.3f}, Neutral={res['3m_probs']['Neutral']:.3f}, Up={res['3m_probs']['Up']:.3f}")
+            print(f"Time: {res['inference_time']*1000:.2f} ms")
+            return res
+        # Warm-up
+        print("\n--- Warm-up Run ---")
+        predictor.predict("Warmup", "WP", "Warmup", {
+            "feat_ret_1m": 0.0, "feat_ret_5m": 0.0, "feat_ret_15m": 0.0,
+            "feat_volatility_60m": 0.0, "feat_num_trades_60m": 0, "feat_volume_60m": 0,
+            "feat_tweet_freq_24h": 0, "feat_time_since_prev_tweet": 0,
+            "feat_btc_ret_60m": 0.0, "feat_btc_ret_24h": 0.0,
+            "feat_fear_greed_index": 50, "feat_btc_dominance": 50, "feat_altseason_index": 0
+        })
+        # 1. FOMO / Strong Uptrend
+        run_scenario(
+            "FOMO / Strong Uptrend",
+            "BREAKING: Major exchange listing confirmed! 🚀 #ToTheMoon",
+            {
+                "feat_ret_1m": 0.02, "feat_ret_5m": 0.05, "feat_ret_15m": 0.08,
+                "feat_volatility_60m": 0.05, "feat_num_trades_60m": 1000, "feat_volume_60m": 2000000,
+                "feat_tweet_freq_24h": 100, "feat_time_since_prev_tweet": 5,
+                "feat_btc_ret_60m": 0.01, "feat_btc_ret_24h": 0.05,
+                "feat_fear_greed_index": 80, "feat_btc_dominance": 45, "feat_altseason_index": 80
+            }
+        )
+        # 2. Panic Dump / Crash (The "Mean Reversion" Test)
+        run_scenario(
+            "Panic Dump / Crash",
+            "URGENT: Security breach detected. Do not interact with contracts.",
+            {
+                "feat_ret_1m": -0.05, "feat_ret_5m": -0.08, "feat_ret_15m": -0.10,
+                "feat_volatility_60m": 0.15, "feat_num_trades_60m": 2000, "feat_volume_60m": 5000000,
+                "feat_tweet_freq_24h": 200, "feat_time_since_prev_tweet": 1,
+                "feat_btc_ret_60m": -0.03, "feat_btc_ret_24h": -0.08,
+                "feat_fear_greed_index": 10, "feat_btc_dominance": 60, "feat_altseason_index": 5
+            }
+        )
+        # 3. Slow Bleed / Bear Market
+        run_scenario(
+            "Slow Bleed / Bear Market",
+            "Weekly development update. Progress is slow but steady.",
+            {
+                "feat_ret_1m": -0.001, "feat_ret_5m": -0.002, "feat_ret_15m": -0.005,
+                "feat_volatility_60m": 0.01, "feat_num_trades_60m": 50, "feat_volume_60m": 100000,
+                "feat_tweet_freq_24h": 10, "feat_time_since_prev_tweet": 60,
+                "feat_btc_ret_60m": -0.001, "feat_btc_ret_24h": -0.01,
+                "feat_fear_greed_index": 30, "feat_btc_dominance": 55, "feat_altseason_index": 10
+            }
+        )
+        # 4. Sideways / Stable
+        run_scenario(
+            "Sideways / Stable",
+            "Just a normal day building. #Crypto",
+            {
+                "feat_ret_1m": 0.000, "feat_ret_5m": 0.001, "feat_ret_15m": -0.001,
+                "feat_volatility_60m": 0.005, "feat_num_trades_60m": 20, "feat_volume_60m": 50000,
+                "feat_tweet_freq_24h": 5, "feat_time_since_prev_tweet": 120,
+                "feat_btc_ret_60m": 0.000, "feat_btc_ret_24h": 0.002,
+                "feat_fear_greed_index": 50, "feat_btc_dominance": 50, "feat_altseason_index": 20
+            }
+        )
+        # 5. Divergence: Good News + Bad Price (Opportunity?)
+        run_scenario(
+            "Divergence: Good News + Bad Price",
+            "Partnership with Google Cloud announced!",
+            {
+                "feat_ret_1m": -0.02, "feat_ret_5m": -0.03, "feat_ret_15m": -0.03,
+                "feat_volatility_60m": 0.04, "feat_num_trades_60m": 150, "feat_volume_60m": 300000,
+                "feat_tweet_freq_24h": 20, "feat_time_since_prev_tweet": 10,
+                "feat_btc_ret_60m": -0.01, "feat_btc_ret_24h": -0.02,
+                "feat_fear_greed_index": 40, "feat_btc_dominance": 52, "feat_altseason_index": 15
+            }
+        )
+    else:
+        print(f"Error: Checkpoint folder '{checkpoint_folder}' not found.")
+        print("Please place this script next to your model folder.")