refactor: 清理测试文件和2019相关代码
- 删除测试脚本: test_2026_classification.py, test_lagllama.py, test_timesfm.py, test_ttm.py - 删除2019相关文件: analyze_2019_*, apply_2019_*, dlinear_2019_*, three_models_2019_* - 添加集成模型相关代码和分析脚本 - 添加N-BEATS模型和CatBoost分类模型
This commit is contained in:
crypto_predictor committed
1 parent
c74f4f6949
commit
1bd9d25f60
33 files changed
+6557
-233
No files matched your search
@@ -0,0 +1,406 @@
|
||||
"""分析DLinear输出概率值的偏向性及其对集成策略的影响"""
|
||||
|
||||
import sys
|
||||
import os
|
||||
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
||||
|
||||
import torch
|
||||
import pickle
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
import pandas_ta as ta
|
||||
from sklearn.preprocessing import StandardScaler
|
||||
from data.fetcher import BinanceDataFetcher
|
||||
|
||||
|
||||
# ==================== 预处理器定义 ====================
|
||||
|
||||
class ClassificationPreprocessor:
|
||||
"""DLinear分类预处理器"""
|
||||
def __init__(self, lookback=168, forecast_horizon=1):
|
||||
self.lookback = lookback
|
||||
self.forecast_horizon = forecast_horizon
|
||||
self.scaler = StandardScaler()
|
||||
self.is_fitted = False
|
||||
|
||||
def add_features(self, df):
|
||||
df = df.copy()
|
||||
df['rsi'] = ta.rsi(df['close'], length=14)
|
||||
macd = ta.macd(df['close'])
|
||||
if macd is not None and not macd.empty:
|
||||
df['macd'] = macd.iloc[:, 0]
|
||||
df['macd_hist'] = macd.iloc[:, 1]
|
||||
df['macd_signal'] = macd.iloc[:, 2]
|
||||
return df
|
||||
|
||||
def create_labels(self, df):
|
||||
df = df.copy()
|
||||
df['future_price'] = df['close'].shift(-self.forecast_horizon)
|
||||
df['future_return'] = (df['future_price'] - df['close']) / df['close']
|
||||
df['label'] = (df['future_return'] > 0).astype(int)
|
||||
return df[:-self.forecast_horizon]
|
||||
|
||||
def preprocess(self, df, fit=True):
|
||||
df = self.add_features(df)
|
||||
df = self.create_labels(df)
|
||||
df = df.dropna().reset_index(drop=True)
|
||||
|
||||
feature_cols = ['open', 'high', 'low', 'close', 'volume',
|
||||
'rsi', 'macd', 'macd_hist', 'macd_signal']
|
||||
|
||||
if fit:
|
||||
scaled_features = self.scaler.fit_transform(df[feature_cols])
|
||||
self.is_fitted = True
|
||||
self.feature_columns = feature_cols
|
||||
else:
|
||||
scaled_features = self.scaler.transform(df[feature_cols])
|
||||
|
||||
X, y = [], []
|
||||
for i in range(len(scaled_features) - self.lookback):
|
||||
X.append(scaled_features[i:i + self.lookback])
|
||||
y.append(df['label'].iloc[i + self.lookback])
|
||||
|
||||
return np.array(X), np.array(y), df
|
||||
|
||||
|
||||
class NBeatsPreprocessor:
|
||||
"""N-BEATS预处理器"""
|
||||
def __init__(self, lookback=168, forecast_horizon=1):
|
||||
self.lookback = lookback
|
||||
self.forecast_horizon = forecast_horizon
|
||||
self.price_scaler = StandardScaler()
|
||||
self.is_fitted = False
|
||||
|
||||
def preprocess(self, df, fit=True):
|
||||
df = df.copy()
|
||||
df['future_price'] = df['close'].shift(-self.forecast_horizon)
|
||||
df = df[:-self.forecast_horizon].reset_index(drop=True)
|
||||
|
||||
prices = df['close'].values.reshape(-1, 1)
|
||||
future_prices = df['future_price'].values.reshape(-1, 1)
|
||||
|
||||
if fit:
|
||||
scaled_prices = self.price_scaler.fit_transform(prices)
|
||||
scaled_future = self.price_scaler.transform(future_prices)
|
||||
self.is_fitted = True
|
||||
else:
|
||||
scaled_prices = self.price_scaler.transform(prices)
|
||||
scaled_future = self.price_scaler.transform(future_prices)
|
||||
|
||||
X, y = [], []
|
||||
for i in range(len(scaled_prices) - self.lookback):
|
||||
X.append(scaled_prices[i:i + self.lookback, 0])
|
||||
y.append(scaled_future[i + self.lookback, 0])
|
||||
|
||||
return np.array(X), np.array(y), df
|
||||
|
||||
def inverse_transform_price(self, scaled_price):
|
||||
return self.price_scaler.inverse_transform(scaled_price.reshape(-1, 1)).flatten()
|
||||
|
||||
|
||||
def main():
|
||||
print("="*80)
|
||||
print("DLinear输出概率值分布分析 - 以0.1阶梯统计")
|
||||
print("="*80)
|
||||
|
||||
# ========== 加载模型 ==========
|
||||
print("\n[1/4] 加载模型...")
|
||||
|
||||
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
|
||||
|
||||
# 加载DLinear
|
||||
from models.dlinear.model import DLinear
|
||||
dlinear_checkpoint = torch.load(
|
||||
'./models/checkpoints/dlinear_classification_2026/best_model.pt',
|
||||
map_location=device
|
||||
)
|
||||
dlinear_model = DLinear(dlinear_checkpoint['config'])
|
||||
dlinear_model.load_state_dict(dlinear_checkpoint['model_state_dict'])
|
||||
dlinear_model = dlinear_model.to(device)
|
||||
dlinear_model.eval()
|
||||
|
||||
with open('./models/checkpoints/dlinear_classification_2026/preprocessor.pkl', 'rb') as f:
|
||||
dlinear_preprocessor = pickle.load(f)
|
||||
|
||||
# 加载N-BEATS
|
||||
from models.nbeats.model import NBeats
|
||||
nbeats_checkpoint = torch.load(
|
||||
'./models/checkpoints/nbeats_btc_1h/best_model.pt',
|
||||
map_location=device
|
||||
)
|
||||
nbeats_model = NBeats(nbeats_checkpoint['config'])
|
||||
nbeats_model.load_state_dict(nbeats_checkpoint['model_state_dict'])
|
||||
nbeats_model = nbeats_model.to(device)
|
||||
nbeats_model.eval()
|
||||
|
||||
with open('./models/checkpoints/nbeats_btc_1h/preprocessor.pkl', 'rb') as f:
|
||||
nbeats_preprocessor = pickle.load(f)
|
||||
|
||||
# 加载TTM-R2
|
||||
from tsfm_public.models.tinytimemixer import TinyTimeMixerForPrediction
|
||||
ttm_model = TinyTimeMixerForPrediction.from_pretrained(
|
||||
"ibm-granite/granite-timeseries-ttm-r2",
|
||||
trust_remote_code=True
|
||||
)
|
||||
ttm_model = ttm_model.to(device)
|
||||
ttm_model.eval()
|
||||
|
||||
print(f"✓ 三个模型已加载")
|
||||
|
||||
# ========== 获取数据 ==========
|
||||
print("\n[2/4] 获取2026年测试数据...")
|
||||
fetcher = BinanceDataFetcher()
|
||||
|
||||
try:
|
||||
df_2026 = fetcher.fetch_ohlcv('BTC', '1h', start_date='2026-01-01',
|
||||
end_date='2026-12-31', use_cache=False)
|
||||
except:
|
||||
df_2026 = fetcher.fetch_ohlcv('BTC', '1h', start_date='2025-12-01', use_cache=False)
|
||||
|
||||
print(f"✓ 测试数据: {len(df_2026)} 行")
|
||||
|
||||
# ========== 预处理 ==========
|
||||
print("\n[3/4] 预处理数据...")
|
||||
|
||||
X_dlinear, y_dlinear, df_dlinear = dlinear_preprocessor.preprocess(df_2026, fit=False)
|
||||
X_nbeats, y_nbeats, df_nbeats = nbeats_preprocessor.preprocess(df_2026, fit=False)
|
||||
|
||||
min_samples = min(len(X_dlinear), len(X_nbeats))
|
||||
X_dlinear = X_dlinear[:min_samples]
|
||||
y_dlinear = y_dlinear[:min_samples]
|
||||
X_nbeats = X_nbeats[:min_samples]
|
||||
y_nbeats = y_nbeats[:min_samples]
|
||||
|
||||
# ========== 预测并获取DLinear概率值 ==========
|
||||
print("\n[4/4] 执行预测并分析DLinear输出概率...")
|
||||
|
||||
# DLinear预测（保留概率值）
|
||||
X_dlinear_tensor = torch.FloatTensor(X_dlinear).to(device)
|
||||
with torch.no_grad():
|
||||
dlinear_logits = dlinear_model(X_dlinear_tensor).cpu().numpy()
|
||||
dlinear_probs = 1 / (1 + np.exp(-dlinear_logits.flatten())) # sigmoid概率
|
||||
dlinear_preds = (dlinear_probs > 0.5).astype(int)
|
||||
|
||||
# N-BEATS预测
|
||||
X_nbeats_tensor = torch.FloatTensor(X_nbeats).to(device)
|
||||
with torch.no_grad():
|
||||
nbeats_pred_scaled = nbeats_model(X_nbeats_tensor).cpu().numpy()
|
||||
|
||||
nbeats_pred_prices = nbeats_preprocessor.inverse_transform_price(nbeats_pred_scaled)
|
||||
current_open_prices = df_nbeats['open'].iloc[168:168+len(nbeats_pred_prices)].values
|
||||
nbeats_preds = (nbeats_pred_prices > current_open_prices).astype(int)
|
||||
|
||||
# TTM-R2预测
|
||||
context_length = 512
|
||||
ttm_preds_list = []
|
||||
all_prices = df_2026['close'].values
|
||||
all_open_prices = df_2026['open'].values
|
||||
|
||||
for i in range(min_samples):
|
||||
start_idx = i
|
||||
end_idx = start_idx + context_length
|
||||
|
||||
if end_idx > len(all_prices):
|
||||
context_data = all_prices[-context_length:].reshape(1, -1, 1)
|
||||
current_open_price = all_open_prices[min(168 + i, len(all_open_prices) - 1)]
|
||||
else:
|
||||
context_data = all_prices[start_idx:end_idx].reshape(1, -1, 1)
|
||||
current_open_price = all_open_prices[end_idx] if end_idx < len(all_open_prices) else all_open_prices[-1]
|
||||
|
||||
context_tensor = torch.FloatTensor(context_data).to(device)
|
||||
|
||||
with torch.no_grad():
|
||||
output = ttm_model(context_tensor)
|
||||
pred_price = output.prediction_outputs.cpu().numpy()[0, 0, 0]
|
||||
|
||||
ttm_preds_list.append(1 if pred_price > current_open_price else 0)
|
||||
|
||||
ttm_preds = np.array(ttm_preds_list)
|
||||
|
||||
# 双模型交集
|
||||
two_model_agree_up = (dlinear_preds == 1) & (nbeats_preds == 1)
|
||||
two_model_agree_down = (dlinear_preds == 0) & (nbeats_preds == 0)
|
||||
two_model_valid = two_model_agree_up | two_model_agree_down
|
||||
|
||||
# 三模型多数投票
|
||||
vote_sum = dlinear_preds + nbeats_preds + ttm_preds
|
||||
three_model_preds = np.where(vote_sum >= 2, 1, 0)
|
||||
|
||||
# 真实标签
|
||||
true_labels = y_dlinear
|
||||
|
||||
print(f"✓ 预测完成，开始分析DLinear概率分布...")
|
||||
|
||||
# ========== 按0.1阶梯分析 ==========
|
||||
print("\n" + "="*80)
|
||||
print("📊 DLinear输出概率分布分析（0.1阶梯）")
|
||||
print("="*80)
|
||||
|
||||
# 定义概率区间
|
||||
prob_bins = np.arange(0, 1.1, 0.1)
|
||||
prob_labels = [f"{i/10:.1f}-{(i+1)/10:.1f}" for i in range(10)]
|
||||
|
||||
# 为每个样本分配到概率区间
|
||||
prob_bin_indices = np.digitize(dlinear_probs, prob_bins) - 1
|
||||
prob_bin_indices = np.clip(prob_bin_indices, 0, 9) # 确保在0-9范围内
|
||||
|
||||
# 初始化统计
|
||||
stats = []
|
||||
|
||||
for i, bin_label in enumerate(prob_labels):
|
||||
mask = prob_bin_indices == i
|
||||
count = mask.sum()
|
||||
|
||||
if count == 0:
|
||||
continue
|
||||
|
||||
# 该区间的真实标签
|
||||
bin_true = true_labels[mask]
|
||||
|
||||
# 该区间DLinear的预测
|
||||
bin_dlinear_pred = dlinear_preds[mask]
|
||||
|
||||
# 该区间在双模型交集中的表现
|
||||
bin_two_model_valid = two_model_valid[mask]
|
||||
bin_two_model_pred = np.full(count, -1)
|
||||
bin_two_model_pred[two_model_agree_up[mask]] = 1
|
||||
bin_two_model_pred[two_model_agree_down[mask]] = 0
|
||||
|
||||
# 该区间在三模型投票中的表现
|
||||
bin_three_model_pred = three_model_preds[mask]
|
||||
|
||||
# 计算统计
|
||||
dlinear_acc = (bin_dlinear_pred == bin_true).mean() * 100 if count > 0 else 0
|
||||
|
||||
two_model_valid_count = bin_two_model_valid.sum()
|
||||
two_model_acc = ((bin_two_model_pred == bin_true) & (bin_two_model_pred != -1)).sum() / two_model_valid_count * 100 if two_model_valid_count > 0 else 0
|
||||
two_model_coverage = two_model_valid_count / count * 100
|
||||
|
||||
three_model_acc = (bin_three_model_pred == bin_true).mean() * 100 if count > 0 else 0
|
||||
|
||||
# 平均概率值
|
||||
avg_prob = dlinear_probs[mask].mean()
|
||||
|
||||
# 预测倾向
|
||||
pred_up = bin_dlinear_pred.mean() * 100
|
||||
|
||||
stats.append({
|
||||
'bin': bin_label,
|
||||
'count': count,
|
||||
'avg_prob': avg_prob,
|
||||
'pred_up_pct': pred_up,
|
||||
'dlinear_acc': dlinear_acc,
|
||||
'two_model_acc': two_model_acc,
|
||||
'two_model_cov': two_model_coverage,
|
||||
'three_model_acc': three_model_acc
|
||||
})
|
||||
|
||||
# 转换为DataFrame
|
||||
stats_df = pd.DataFrame(stats)
|
||||
|
||||
# 显示统计表
|
||||
print("\n概率区间分析：")
|
||||
print(f"{'区间':<10} {'样本数':>6} {'平均概率':>8} {'预测↑%':>7} "
|
||||
f"{'DL准确率':>9} {'双模准确率':>10} {'双模覆盖':>8} {'三模准确率':>10}")
|
||||
print("-" * 95)
|
||||
|
||||
for _, row in stats_df.iterrows():
|
||||
# 根据准确率添加标记
|
||||
dl_mark = "🔥" if row['dlinear_acc'] >= 70 else ("✅" if row['dlinear_acc'] >= 60 else ("⚠️" if row['dlinear_acc'] < 45 else ""))
|
||||
two_mark = "🔥" if row['two_model_acc'] >= 70 else ("✅" if row['two_model_acc'] >= 60 else ("⚠️" if row['two_model_acc'] < 45 else ""))
|
||||
three_mark = "🔥" if row['three_model_acc'] >= 70 else ("✅" if row['three_model_acc'] >= 60 else ("⚠️" if row['three_model_acc'] < 45 else ""))
|
||||
|
||||
print(f"{row['bin']:<10} {row['count']:>6} {row['avg_prob']:>8.3f} {row['pred_up_pct']:>6.1f}% "
|
||||
f"{row['dlinear_acc']:>8.1f}{dl_mark} "
|
||||
f"{row['two_model_acc']:>9.1f}{two_mark} "
|
||||
f"{row['two_model_cov']:>7.1f}% "
|
||||
f"{row['three_model_acc']:>9.1f}{three_mark}")
|
||||
|
||||
# ========== 关键发现 ==========
|
||||
print("\n" + "="*80)
|
||||
print("🔍 关键发现")
|
||||
print("="*80)
|
||||
|
||||
# 极端概率区间（高置信度）
|
||||
high_conf_up = stats_df[stats_df['avg_prob'] >= 0.65]
|
||||
high_conf_down = stats_df[stats_df['avg_prob'] <= 0.35]
|
||||
|
||||
print(f"\n📈 高置信度看涨区间 (概率≥0.65):")
|
||||
if len(high_conf_up) > 0:
|
||||
total_samples = high_conf_up['count'].sum()
|
||||
avg_dl_acc = (high_conf_up['dlinear_acc'] * high_conf_up['count']).sum() / total_samples
|
||||
avg_two_acc = (high_conf_up['two_model_acc'] * high_conf_up['count']).sum() / total_samples
|
||||
avg_three_acc = (high_conf_up['three_model_acc'] * high_conf_up['count']).sum() / total_samples
|
||||
|
||||
print(f" 样本数: {total_samples} ({total_samples/len(dlinear_probs)*100:.1f}%)")
|
||||
print(f" DLinear平均准确率: {avg_dl_acc:.2f}%")
|
||||
print(f" 双模型平均准确率: {avg_two_acc:.2f}%")
|
||||
print(f" 三模型平均准确率: {avg_three_acc:.2f}%")
|
||||
else:
|
||||
print(" 无数据")
|
||||
|
||||
print(f"\n📉 高置信度看跌区间 (概率≤0.35):")
|
||||
if len(high_conf_down) > 0:
|
||||
total_samples = high_conf_down['count'].sum()
|
||||
avg_dl_acc = (high_conf_down['dlinear_acc'] * high_conf_down['count']).sum() / total_samples
|
||||
avg_two_acc = (high_conf_down['two_model_acc'] * high_conf_down['count']).sum() / total_samples
|
||||
avg_three_acc = (high_conf_down['three_model_acc'] * high_conf_down['count']).sum() / total_samples
|
||||
|
||||
print(f" 样本数: {total_samples} ({total_samples/len(dlinear_probs)*100:.1f}%)")
|
||||
print(f" DLinear平均准确率: {avg_dl_acc:.2f}%")
|
||||
print(f" 双模型平均准确率: {avg_two_acc:.2f}%")
|
||||
print(f" 三模型平均准确率: {avg_three_acc:.2f}%")
|
||||
else:
|
||||
print(" 无数据")
|
||||
|
||||
# 中等置信度区间
|
||||
mid_conf = stats_df[(stats_df['avg_prob'] > 0.45) & (stats_df['avg_prob'] < 0.55)]
|
||||
|
||||
print(f"\n⚖️ 中等置信度区间 (0.45-0.55):")
|
||||
if len(mid_conf) > 0:
|
||||
total_samples = mid_conf['count'].sum()
|
||||
avg_dl_acc = (mid_conf['dlinear_acc'] * mid_conf['count']).sum() / total_samples
|
||||
avg_two_acc = (mid_conf['two_model_acc'] * mid_conf['count']).sum() / total_samples if total_samples > 0 else 0
|
||||
avg_three_acc = (mid_conf['three_model_acc'] * mid_conf['count']).sum() / total_samples
|
||||
|
||||
print(f" 样本数: {total_samples} ({total_samples/len(dlinear_probs)*100:.1f}%)")
|
||||
print(f" DLinear平均准确率: {avg_dl_acc:.2f}%")
|
||||
print(f" 双模型平均准确率: {avg_two_acc:.2f}%")
|
||||
print(f" 三模型平均准确率: {avg_three_acc:.2f}%")
|
||||
print(f" ⚠️ 建议: 低置信度区间，考虑弃权或减仓")
|
||||
else:
|
||||
print(" 无数据")
|
||||
|
||||
# 找出表现最好的概率区间
|
||||
best_bin = stats_df.loc[stats_df['dlinear_acc'].idxmax()]
|
||||
worst_bin = stats_df.loc[stats_df['dlinear_acc'].idxmin()]
|
||||
|
||||
print(f"\n🏆 DLinear表现最佳区间:")
|
||||
print(f" 区间: {best_bin['bin']}")
|
||||
print(f" 准确率: {best_bin['dlinear_acc']:.1f}%")
|
||||
print(f" 样本数: {best_bin['count']}")
|
||||
|
||||
print(f"\n💀 DLinear表现最差区间:")
|
||||
print(f" 区间: {worst_bin['bin']}")
|
||||
print(f" 准确率: {worst_bin['dlinear_acc']:.1f}%")
|
||||
print(f" 样本数: {worst_bin['count']}")
|
||||
|
||||
# ========== 概率分布可视化 ==========
|
||||
print("\n" + "="*80)
|
||||
print("📊 DLinear概率分布直方图")
|
||||
print("="*80)
|
||||
|
||||
print("\n样本数分布:")
|
||||
max_count = stats_df['count'].max()
|
||||
for _, row in stats_df.iterrows():
|
||||
bar_length = int(row['count'] / max_count * 40)
|
||||
bar = "█" * bar_length
|
||||
print(f"{row['bin']:<10} {bar} {row['count']}")
|
||||
|
||||
print("\n" + "="*80)
|
||||
print("✓ DLinear概率分布分析完成!")
|
||||
print("="*80)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in new issue
Block a user