diff --git a/crypto/gogo/data/data_utils.py b/crypto/gogo/data/data_utils.py
index 32474c3..0dd142e 100644
--- a/crypto/gogo/data/data_utils.py
+++ b/crypto/gogo/data/data_utils.py
@@ -121,6 +121,44 @@ def create_padding_mask(seq, pad_token=0):
     """
     return (seq == pad_token).all(dim=-1).unsqueeze(0)
 
+def get_aligned_candle_with_index(candles_list, base_ts):
+    """
+    Find the candle from candles_list that is closest to (and <=) base_ts.
+    Returns: (index, candle)
+    """
+    aligned_index = None
+    aligned_candle = None
+    for i in range(len(candles_list)):
+        if candles_list[i]["timestamp"] <= base_ts:
+            aligned_index = i
+            aligned_candle = candles_list[i]
+        else:
+            break
+    return aligned_index, aligned_candle
+
+def get_features_for_tf(candles_list, aligned_index, period=10):
+    """
+    Extract features from the candle at aligned_index.
+    If aligned_index is None, return a zeroed feature vector.
+    """
+    if aligned_index is None:
+        return [0.0] * 7  # return zeroed feature vector
+    candle = candles_list[aligned_index]
+    # Simple features: open, high, low, close, volume, and two EMAs.
+    close_prices = [c["close"] for c in candles_list[:aligned_index+1]]
+    ema_short = calculate_ema(candles_list[:aligned_index+1], period=period)[-1]
+    ema_long = calculate_ema(candles_list[:aligned_index+1], period=period*2)[-1]
+    features = [
+        candle["open"],
+        candle["high"],
+        candle["low"],
+        candle["close"],
+        candle["volume"],
+        ema_short,
+        ema_long
+    ]
+    return features
+
 # Example usage (within a larger training loop):
 if __name__ == '__main__':
     # Dummy data for demonstration
@@ -155,4 +193,14 @@ if __name__ == '__main__':
     mask = create_mask(seq_len)
     print("\nMask:\n", mask)
     padding_mask = create_padding_mask(torch.tensor(candle_features))
-    print(f"\nPadding mask: {padding_mask}")
\ No newline at end of file
+    print(f"\nPadding mask: {padding_mask}")
+
+    # Example usage of the new functions
+    index, candle = get_aligned_candle_with_index(candles_data, 1678886570000)
+    if candle:
+        print(f"\nAligned candle: {candle}")
+    else:
+        print("\nNo aligned candle found.")
+
+    features = get_features_for_tf(candles_data, index)
+    print(f"\nFeatures for timeframe: {features}")
diff --git a/crypto/gogo/data/live_data.py b/crypto/gogo/data/live_data.py
index 6b843d6..fa6a76f 100644
--- a/crypto/gogo/data/live_data.py
+++ b/crypto/gogo/data/live_data.py
@@ -7,7 +7,7 @@ from collections import deque
 
 import ccxt.async_support as ccxt
 from dotenv import load_dotenv
-
+import platform
 
 class LiveDataManager:
     def __init__(self, symbol, exchange_name='mexc', window_size=120):
@@ -20,6 +20,7 @@ class LiveDataManager:
         self.last_candle_time = None
         self.exchange = self._initialize_exchange()
         self.lock = asyncio.Lock() # Lock to prevent race conditions
+        self.is_windows = platform.system() == 'Windows'
 
     def _initialize_exchange(self):
         exchange_class = getattr(ccxt, self.exchange_name)
@@ -41,15 +42,23 @@ class LiveDataManager:
         print(f"Fetching initial candles for {self.symbol}...")
         now = int(time.time() * 1000)
         since = now - self.window_size * 60 * 1000
-        try:
-            candles = await self.exchange.fetch_ohlcv(self.symbol, '1m', since=since, limit=self.window_size)
-            for candle in candles:
-                self.candles.append(self._format_candle(candle))
-            if candles:
-                self.last_candle_time = candles[-1][0]
-            print(f"Fetched {len(candles)} initial candles.")
-        except Exception as e:
-            print(f"Error fetching initial candles: {e}")
+        retries = 3
+        for attempt in range(retries):
+            try:
+                candles = await self.exchange.fetch_ohlcv(self.symbol, '1m', since=since, limit=self.window_size)
+                for candle in candles:
+                    self.candles.append(self._format_candle(candle))
+                if candles:
+                    self.last_candle_time = candles[-1][0]
+                print(f"Fetched {len(candles)} initial candles.")
+                return  # Exit the function if successful
+            except Exception as e:
+                print(f"Attempt {attempt + 1} failed: {e}")
+                if self.is_windows and "aiodns needs a SelectorEventLoop" in str(e):
+                    print("aiodns issue detected on Windows.  This is a known problem with aiodns and ccxt on Windows.")
+                if attempt < retries - 1:
+                    await asyncio.sleep(5)  # Wait before retrying
+        print("Failed to fetch initial candles after multiple retries.")
 
     def _format_candle(self, candle_data):
         return {
@@ -112,16 +121,23 @@ class LiveDataManager:
     async def fetch_and_process_ticks(self):
         async with self.lock:
             since = None if not self.ticks else self.ticks[-1]['timestamp']
-            try:
-                # Use fetch_trades (or appropriate method for your exchange) for live ticks.
-                ticks = await self.exchange.fetch_trades(self.symbol, since=since)
-                for tick in ticks:
-                    formatted_tick = self._format_tick(tick)
-                    if formatted_tick:  # Add the check here
-                        self.ticks.append(formatted_tick)
-                        await self._update_candle(formatted_tick)
-            except Exception as e:
-                print(f"Error fetching ticks: {e}")
+            retries = 3
+            for attempt in range(retries):
+                try:
+                    # Use fetch_trades (or appropriate method for your exchange) for live ticks.
+                    ticks = await self.exchange.fetch_trades(self.symbol, since=since)
+                    for tick in ticks:
+                        formatted_tick = self._format_tick(tick)
+                        if formatted_tick:  # Add the check here
+                            self.ticks.append(formatted_tick)
+                            await self._update_candle(formatted_tick)
+                    break  # Exit the retry loop if successful
+                except Exception as e:
+                    print(f"Error fetching ticks (attempt {attempt + 1}): {e}")
+                    if self.is_windows and "aiodns needs a SelectorEventLoop" in str(e):
+                        print("aiodns issue detected on Windows.  This is a known problem with aiodns and ccxt on Windows.")
+                    if attempt < retries - 1:
+                        await asyncio.sleep(5)  # Wait before retrying
 
     async def get_data(self):
         async with self.lock:
diff --git a/crypto/gogo/main.py b/crypto/gogo/main.py
index c819c09..60e9124 100644
--- a/crypto/gogo/main.py
+++ b/crypto/gogo/main.py
@@ -8,6 +8,15 @@ from data.live_data import LiveDataManager
 from model.transformer import Transformer
 from training.train import train
 from data.data_utils import preprocess_data  # Import preprocess_data
+import ccxt.async_support as ccxt
+import time
+import os
+import numpy as np
+import matplotlib.pyplot as plt
+from model.trading_model import TradingModel
+from training.rl_agent import ContinuousRLAgent, ReplayBuffer
+from training.train_historical import train_on_historical_data, load_best_checkpoint, save_candles_cache, CACHE_FILE, BEST_DIR
+from data.data_utils import get_aligned_candle_with_index, get_features_for_tf
 
 async def main():
     symbol = 'BTC/USDT'
@@ -43,5 +52,245 @@ async def main():
         print("Training stopped manually.")
     finally:
         await data_manager.close()
+
+# -------------------------------------
+# Main Asynchronous Function for Training & Charting
+# -------------------------------------
+async def main_backtest():
+    symbol = 'BTC/USDT'
+    # Define timeframes: we'll use 5 different ones.
+    timeframes = ["1m", "5m", "15m", "1h", "1d"]
+    now = int(time.time() * 1000)
+    # Use the base timeframe period of 1500 candles. For 1m, that is 1500 minutes.
+    period_ms = 1500 * 60 * 1000
+    since = now - period_ms
+    end_time = now
+
+    # Initialize exchange using MEXC (or your preferred exchange).
+    mexc_api_key = os.environ.get('MEXC_API_KEY', 'YOUR_API_KEY')
+    mexc_api_secret = os.environ.get('MEXC_API_SECRET', 'YOUR_SECRET_KEY')
+    exchange = ccxt.mexc({
+        'apiKey': mexc_api_key,
+        'secret': mexc_api_secret,
+        'enableRateLimit': True,
+    })
+
+    candles_dict = {}
+    for tf in timeframes:
+        print(f"Fetching historical data for timeframe {tf}...")
+        candles = await fetch_historical_data(exchange, symbol, tf, since, end_time, batch_size=500)
+        candles_dict[tf] = candles
+
+    # Optionally, save the multi-timeframe cache.
+    save_candles_cache(CACHE_FILE, candles_dict)
+
+    # Create the backtest environment using multi-timeframe data.
+    env = BacktestEnvironment(candles_dict, base_tf="1m", timeframes=timeframes)
+    
+    # Neural Network dimensions: each timeframe produces 7 features.
+    input_dim = len(timeframes) * 7  # 7 features * 5 timeframes = 35.
+    hidden_dim = 128
+    output_dim = 3  # Actions: SELL, HOLD, BUY.
+
+    model = TradingModel(input_dim, hidden_dim, output_dim)
+    optimizer = optim.Adam(model.parameters(), lr=1e-4)
+    replay_buffer = ReplayBuffer(capacity=10000)
+    rl_agent = ContinuousRLAgent(model, optimizer, replay_buffer, batch_size=32, gamma=0.99)
+
+    # Load best checkpoint if available.
+    load_best_checkpoint(model, BEST_DIR)
+
+    # Train the agent over the historical period.
+    num_epochs = 10  # Adjust as needed.
+    train_on_historical_data(env, rl_agent, num_epochs=num_epochs, epsilon=0.1)
+
+    # Run a final simulation (without exploration) to record trade history.
+    state = env.reset(clear_trade_history=True)
+    done = False
+    cumulative_reward = 0.0
+    while not done:
+        action = rl_agent.act(state, epsilon=0.0)
+        state, reward, next_state, done = env.step(action)
+        cumulative_reward += reward
+        state = next_state
+    print("Final simulation cumulative profit:", cumulative_reward)
+
+    # Evaluate trade performance.
+    trades = env.trade_history
+    num_trades = len(trades)
+    num_wins = sum(1 for trade in trades if trade["pnl"] > 0)
+    win_rate = (num_wins / num_trades * 100) if num_trades > 0 else 0.0
+    total_profit = sum(trade["pnl"] for trade in trades)
+    print(f"Total trades: {num_trades}, Wins: {num_wins}, Win rate: {win_rate:.2f}%, Total Profit: {total_profit:.4f}")
+
+    # Plot chart with buy/sell markers on the base timeframe ("1m").
+    plot_trade_history(candles_dict["1m"], trades)
+    
+    await exchange.close()
+
+# -------------------------------------
+# Historical Data Fetching Function (for a given timeframe)
+# -------------------------------------
+async def fetch_historical_data(exchange, symbol, timeframe, since, end_time, batch_size=500):
+    candles = []
+    since_ms = since
+    while True:
+        try:
+            batch = await exchange.fetch_ohlcv(symbol, timeframe=timeframe, since=since_ms, limit=batch_size)
+        except Exception as e:
+            print(f"Error fetching historical data for {timeframe}:", e)
+            break
+        if not batch:
+            break
+        for c in batch:
+            candle_dict = {
+                'timestamp': c[0],
+                'open': c[1],
+                'high': c[2],
+                'low': c[3],
+                'close': c[4],
+                'volume': c[5]
+            }
+            candles.append(candle_dict)
+        last_timestamp = batch[-1][0]
+        if last_timestamp >= end_time:
+            break
+        since_ms = last_timestamp + 1
+    print(f"Fetched {len(candles)} candles for timeframe {timeframe}.")
+    return candles
+
+# -------------------------------------
+# Backtest Environment with Multi-Timeframe State
+# -------------------------------------
+class BacktestEnvironment:
+    def __init__(self, candles_dict, base_tf="1m", timeframes=None):
+        self.candles_dict = candles_dict  # dict of timeframe: candles_list
+        self.base_tf = base_tf
+        if timeframes is None:
+            self.timeframes = [base_tf]  # fallback to single timeframe
+        else:
+            self.timeframes = timeframes
+        self.trade_history = []   # record of closed trades
+        self.current_index = 0    # index on base_tf candles
+        self.position = None      # active position record
+
+    def reset(self, clear_trade_history=True):
+        self.current_index = 0
+        self.position = None
+        if clear_trade_history:
+            self.trade_history = []
+        return self.get_state(self.current_index)
+
+    def get_state(self, index):
+        """Construct the state as the concatenated features of all timeframes.
+        For each timeframe, find the aligned candle for the base timeframe’s timestamp."""
+        state_features = []
+        base_candle = self.candles_dict[self.base_tf][index]
+        base_ts = base_candle["timestamp"]
+        for tf in self.timeframes:
+            candles_list = self.candles_dict[tf]
+            # Get the candle from this timeframe that is closest to (and <=) base_ts.
+            aligned_index, _ = get_aligned_candle_with_index(candles_list, base_ts)
+            features = get_features_for_tf(candles_list, aligned_index, period=10)
+            state_features.extend(features)
+        return np.array(state_features, dtype=np.float32)
+
+    def step(self, action):
+        """
+        Simulate a trading step based on the base timeframe.
+          - If not in a position and action is BUY (2), record entry at next candle's open.
+          - If in a position and action is SELL (0), record exit at next candle's open, computing PnL.
+        Returns: (current_state, reward, next_state, done)
+        """
+        base_candles = self.candles_dict[self.base_tf]
+        if self.current_index >= len(base_candles) - 1:
+            return self.get_state(self.current_index), 0.0, None, True
+
+        current_state = self.get_state(self.current_index)
+        next_index = self.current_index + 1
+        next_state = self.get_state(next_index)
+        current_candle = base_candles[self.current_index]
+        next_candle = base_candles[next_index]
+        reward = 0.0
+
+        # Action mapping: 0 -> SELL, 1 -> HOLD, 2 -> BUY.
+        if self.position is None:
+            if action == 2:  # BUY signal: enter position at next candle's open.
+                entry_price = next_candle["open"]
+                self.position = {"entry_price": entry_price, "entry_index": self.current_index}
+        else:
+            if action == 0:  # SELL signal: close position at next candle's open.
+                exit_price = next_candle["open"]
+                reward = exit_price - self.position["entry_price"]
+                trade = {
+                    "entry_index": self.position["entry_index"],
+                    "entry_price": self.position["entry_price"],
+                    "exit_index": next_index,
+                    "exit_price": exit_price,
+                    "pnl": reward
+                }
+                self.trade_history.append(trade)
+                self.position = None
+
+        self.current_index = next_index
+        done = (self.current_index >= len(base_candles) - 1)
+        return current_state, reward, next_state, done
+
+# -------------------------------------
+# Chart Plotting: Trade History & PnL
+# -------------------------------------
+def plot_trade_history(candles, trade_history):
+    close_prices = [candle["close"] for candle in candles]
+    x = list(range(len(close_prices)))
+    plt.figure(figsize=(12, 6))
+    plt.plot(x, close_prices, label="Close Price", color="black", linewidth=1)
+
+    # Use these flags so that the label "BUY" or "SELL" is only shown once in the legend.
+    buy_label_added = False
+    sell_label_added = False
+
+    for trade in trade_history:
+        in_idx = trade["entry_index"]
+        out_idx = trade["exit_index"]
+        in_price = trade["entry_price"]
+        out_price = trade["exit_price"]
+        pnl = trade["pnl"]
+
+        # Plot BUY marker ("IN")
+        if not buy_label_added:
+            plt.plot(in_idx, in_price, marker="^", color="green", markersize=10, label="BUY (IN)")
+            buy_label_added = True
+        else:
+            plt.plot(in_idx, in_price, marker="^", color="green", markersize=10)
+        plt.text(in_idx, in_price, " IN", color="green", fontsize=8, verticalalignment="bottom")
+
+        # Plot SELL marker ("OUT")
+        if not sell_label_added:
+            plt.plot(out_idx, out_price, marker="v", color="red", markersize=10, label="SELL (OUT)")
+            sell_label_added = True
+        else:
+            plt.plot(out_idx, out_price, marker="v", color="red", markersize=10)
+        plt.text(out_idx, out_price, " OUT", color="red", fontsize=8, verticalalignment="top")
+        
+        # Annotate the PnL near the SELL marker.
+        plt.text(out_idx, out_price, f" {pnl:+.2f}", color="blue", fontsize=8, verticalalignment="bottom")
+        
+        # Choose line color based on profitability.
+        if pnl > 0:
+            line_color = "green"
+        elif pnl < 0:
+            line_color = "red"
+        else:
+            line_color = "gray"
+        # Draw a dotted line between the buy and sell points.
+        plt.plot([in_idx, out_idx], [in_price, out_price], linestyle="dotted", color=line_color)
+
+    plt.title("Trade History with PnL")
+    plt.xlabel("Base Candle Index (1m)")
+    plt.ylabel("Price")
+    plt.legend()
+    plt.grid(True)
+    plt.show()
+
 if __name__ == '__main__':
-    asyncio.run(main())
\ No newline at end of file
+    asyncio.run(main_backtest())
diff --git a/crypto/gogo/model/trading_model.py b/crypto/gogo/model/trading_model.py
new file mode 100644
index 0000000..2357c58
--- /dev/null
+++ b/crypto/gogo/model/trading_model.py
@@ -0,0 +1,17 @@
+# model/trading_model.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class TradingModel(nn.Module):
+    def __init__(self, input_dim, hidden_dim, output_dim):
+        super(TradingModel, self).__init__()
+        self.fc1 = nn.Linear(input_dim, hidden_dim)
+        self.fc2 = nn.Linear(hidden_dim, hidden_dim)
+        self.fc3 = nn.Linear(hidden_dim, output_dim)
+
+    def forward(self, x):
+        x = F.relu(self.fc1(x))
+        x = F.relu(self.fc2(x))
+        x = self.fc3(x)
+        return x
diff --git a/crypto/gogo/training/rl_agent.py b/crypto/gogo/training/rl_agent.py
new file mode 100644
index 0000000..eca4bca
--- /dev/null
+++ b/crypto/gogo/training/rl_agent.py
@@ -0,0 +1,49 @@
+# training/rl_agent.py
+import random
+from collections import deque
+import numpy as np
+import torch
+
+class ContinuousRLAgent:
+    def __init__(self, model, optimizer, replay_buffer, batch_size=32, gamma=0.99):
+        self.model = model
+        self.optimizer = optimizer
+        self.replay_buffer = replay_buffer
+        self.batch_size = batch_size
+        self.gamma = gamma
+
+    def act(self, state, epsilon=0.0):
+        """
+        Select an action based on the state, using an epsilon-greedy policy.
+        """
+        if random.random() < epsilon:
+            # Exploration: choose a random action.
+            action = np.random.choice([0, 1, 2])  # SELL, HOLD, BUY
+        else:
+            # Exploitation: choose the action with the highest Q-value.
+            with torch.no_grad():
+                state_tensor = torch.tensor(state, dtype=torch.float32).unsqueeze(0)
+                q_values = self.model(state_tensor)
+                action = torch.argmax(q_values).item()
+        return action
+
+class ReplayBuffer:
+    def __init__(self, capacity):
+        self.buffer = deque(maxlen=capacity)
+
+    def push(self, state, action, reward, next_state, done):
+        """
+        Store an experience tuple into the replay buffer.
+        """
+        self.buffer.append((state, action, reward, next_state, done))
+
+    def sample(self, batch_size):
+        """
+        Randomly sample a batch of experiences from the replay buffer.
+        """
+        batch = random.sample(self.buffer, batch_size)
+        states, actions, rewards, next_states, dones = zip(*batch)
+        return states, actions, rewards, next_states, dones
+
+    def __len__(self):
+        return len(self.buffer)
diff --git a/crypto/gogo/training/train_historical.py b/crypto/gogo/training/train_historical.py
new file mode 100644
index 0000000..aacf17c
--- /dev/null
+++ b/crypto/gogo/training/train_historical.py
@@ -0,0 +1,194 @@
+# training/train_historical.py
+import os
+import json
+import time
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from model.trading_model import TradingModel
+from data.data_utils import get_aligned_candle_with_index, get_features_for_tf
+
+# --- Directories for saving models ---
+LAST_DIR = os.path.join("models", "last")
+BEST_DIR = os.path.join("models", "best")
+os.makedirs(LAST_DIR, exist_ok=True)
+os.makedirs(BEST_DIR, exist_ok=True)
+
+# --- File for saving candles cache ---
+CACHE_FILE = "candles_cache.json"
+
+# -------------------------------------
+# Checkpoint Functions (same as before)
+# -------------------------------------
+def maintain_checkpoint_directory(directory, max_files=10):
+    files = os.listdir(directory)
+    if len(files) > max_files:
+        full_paths = [os.path.join(directory, f) for f in files]
+        full_paths.sort(key=lambda x: os.path.getmtime(x))
+        for f in full_paths[: len(files) - max_files]:
+            os.remove(f)
+
+def get_best_models(directory):
+    best_files = []
+    for file in os.listdir(directory):
+        parts = file.split("_")
+        try:
+            r = float(parts[1])
+            best_files.append((r, file))
+        except Exception:
+            continue
+    return best_files
+
+def save_checkpoint(model, epoch, total_loss, last_dir=LAST_DIR, best_dir=BEST_DIR):
+    timestamp = time.strftime("%Y%m%d_%H%M%S")
+    last_filename = f"model_last_epoch_{epoch}_{timestamp}.pt"
+    last_path = os.path.join(last_dir, last_filename)
+    torch.save({
+         "epoch": epoch,
+         "total_loss": total_loss,
+         "model_state_dict": model.state_dict()
+    }, last_path)
+    maintain_checkpoint_directory(last_dir, max_files=10)
+
+    best_models = get_best_models(best_dir)
+    add_to_best = False
+    if len(best_models) < 10:
+        add_to_best = True
+    else:
+        min_loss, min_file = min(best_models, key=lambda x: x[0])
+        if total_loss < min_loss:
+            add_to_best = True
+            os.remove(os.path.join(best_dir, min_file))
+    if add_to_best:
+        best_filename = f"best_{total_loss:.4f}_epoch_{epoch}_{timestamp}.pt"
+        best_path = os.path.join(best_dir, best_filename)
+        torch.save({
+            "epoch": epoch,
+            "total_loss": total_loss,
+            "model_state_dict": model.state_dict()
+        }, best_path)
+        maintain_checkpoint_directory(best_dir, max_files=10)
+    print(f"Saved checkpoint for epoch {epoch} with loss {total_loss:.4f}")
+
+def load_best_checkpoint(model, best_dir=BEST_DIR):
+    best_models = get_best_models(best_dir)
+    if not best_models:
+        return None
+    best_loss, best_file = min(best_models, key=lambda x: x[0]) #changed to min to represent the loss
+    path = os.path.join(best_dir, best_file)
+    print(f"Loading best model from checkpoint: {best_file} with loss {best_loss:.4f}")
+    checkpoint = torch.load(path)
+    model.load_state_dict(checkpoint["model_state_dict"])
+    return checkpoint
+
+# -------------------------------------
+# Training Loop on Historical Data
+# -------------------------------------
+def train_on_historical_data(env, rl_agent, num_epochs=10, epsilon=0.1):
+    """
+    Train the RL agent on historical data using the backtest environment.
+    """
+    model = rl_agent.model
+    optimizer = rl_agent.optimizer
+    replay_buffer = rl_agent.replay_buffer
+    batch_size = rl_agent.batch_size
+    gamma = rl_agent.gamma
+
+    model.train()
+    criterion = nn.MSELoss()  # or another suitable loss
+
+    for epoch in range(num_epochs):
+        state = env.reset()
+        done = False
+        total_reward = 0
+        total_loss = 0
+
+        while not done:
+            # Agent takes action (with exploration).
+            action = rl_agent.act(state, epsilon=epsilon)
+            current_state, reward, next_state, done = env.step(action)
+            total_reward += reward
+
+            # Store experience in replay buffer.
+            replay_buffer.push(state, action, reward, next_state, done)
+
+            # Train on a batch from the replay buffer.
+            if len(replay_buffer) > batch_size:
+                # Sample a batch from the replay buffer.
+                states, actions, rewards, next_states, dones = replay_buffer.sample(batch_size)
+
+                # Convert data to PyTorch tensors.
+                states = torch.tensor(states, dtype=torch.float32)
+                actions = torch.tensor(actions, dtype=torch.float32)
+                rewards = torch.tensor(rewards, dtype=torch.float32)
+                next_states = torch.tensor(next_states, dtype=torch.float32)
+                dones = torch.tensor(dones, dtype=torch.float32)
+
+                # Compute Q-values for current states.
+                q_values = model(states)
+                
+                # Compute Q-values for next states.
+                next_q_values = model(next_states)
+
+                # Compute the TD target.
+                td_target = rewards + gamma * torch.max(next_q_values, dim=1)[0] * (1 - dones)
+
+                # Compute the loss.
+                loss = criterion(q_values.gather(1, actions.long().unsqueeze(1)).squeeze(), td_target)
+
+                # Optimize the model.
+                optimizer.zero_grad()
+                loss.backward()
+                optimizer.step()
+
+                total_loss += loss.item()
+
+            # Move to the next state.
+            state = next_state
+
+        print(f"Epoch {epoch + 1}/{num_epochs}, Total Reward: {total_reward:.4f}, Loss: {total_loss:.4f}")
+        save_checkpoint(model, epoch, total_loss, LAST_DIR, BEST_DIR)
+
+# -------------------------------------
+# Caching Functions (for candles data)
+# -------------------------------------
+def save_candles_cache(filename, candles_dict):
+    """
+    Save the candles data to a JSON file.
+    """
+    # Convert numpy arrays to lists for JSON serialization.
+    serializable_candles_dict = {}
+    for timeframe, candles in candles_dict.items():
+        serializable_candles = []
+        for candle in candles:
+            serializable_candle = {
+                'timestamp': candle['timestamp'],
+                'open': candle['open'],
+                'high': candle['high'],
+                'low': candle['low'],
+                'close': candle['close'],
+                'volume': candle['volume']
+            }
+            serializable_candles.append(serializable_candle)
+        serializable_candles_dict[timeframe] = serializable_candles
+
+    with open(filename, 'w') as f:
+        json.dump(serializable_candles_dict, f)
+
+def load_candles_cache(filename):
+    """
+    Load the candles data from a JSON file.
+    """
+    with open(filename, 'r') as f:
+        candles_dict = json.load(f)
+
+    # Convert lists back to numpy arrays.
+    for timeframe, candles in candles_dict.items():
+        for candle in candles:
+            candle['open'] = float(candle['open'])
+            candle['high'] = float(candle['high'])
+            candle['low'] = float(candle['low'])
+            candle['close'] = float(candle['close'])
+            candle['volume'] = float(candle['volume'])
+    return candles_dict