317 lines
13 KiB
Python
317 lines
13 KiB
Python
"""
|
||
AI策略快速优化 — 只用LightGBM，多时间框架特征，扫描参数
|
||
|
||
优化点 vs v1:
|
||
1. 63个特征(加5m/15m多时间框架)
|
||
2. 更强LightGBM参数(更多树+更深)
|
||
3. 扫描: 概率阈值/止损/止盈/前瞻期/持仓时间
|
||
4. 要求多空概率差距>3%才开仓(减少弱信号)
|
||
5. 动态ATR止盈选项
|
||
|
||
固定: 100U保证金, 100x杠杆, 10,000U名义, 90%返佣
|
||
"""
|
||
import datetime, sqlite3, time as _time
|
||
import numpy as np
|
||
import pandas as pd
|
||
import lightgbm as lgb
|
||
import warnings
|
||
from pathlib import Path
|
||
from collections import defaultdict
|
||
warnings.filterwarnings('ignore')
|
||
|
||
def load_data():
|
||
db = Path(__file__).parent.parent / 'models' / 'database.db'
|
||
s = int(datetime.datetime(2025,1,1).timestamp())*1000
|
||
e = int(datetime.datetime(2026,1,1).timestamp())*1000
|
||
conn = sqlite3.connect(str(db))
|
||
df = pd.read_sql_query(
|
||
f"SELECT id as ts,open,high,low,close FROM bitmart_eth_1m WHERE id>={s} AND id<{e} ORDER BY id", conn)
|
||
conn.close()
|
||
df['datetime'] = pd.to_datetime(df['ts'], unit='ms')
|
||
df.set_index('datetime', inplace=True)
|
||
return df
|
||
|
||
def add_features(df):
|
||
c=df['close']; h=df['high']; l=df['low']; o=df['open']
|
||
|
||
for p in [5,8,13,21,50,120]:
|
||
df[f'ema_{p}'] = c.ewm(span=p, adjust=False).mean()
|
||
df['ema_fast_slow'] = (df['ema_8']-df['ema_21'])/c
|
||
df['ema_slow_big'] = (df['ema_21']-df['ema_120'])/c
|
||
df['price_vs_ema120'] = (c-df['ema_120'])/c
|
||
df['price_vs_ema50'] = (c-df['ema_50'])/c
|
||
df['ema8_slope'] = df['ema_8'].pct_change(5)
|
||
df['ema21_slope'] = df['ema_21'].pct_change(5)
|
||
df['ema120_slope'] = df['ema_120'].pct_change(20)
|
||
df['triple_bull'] = ((df['ema_8']>df['ema_21'])&(df['ema_21']>df['ema_120'])).astype(float)
|
||
df['triple_bear'] = ((df['ema_8']<df['ema_21'])&(df['ema_21']<df['ema_120'])).astype(float)
|
||
|
||
delta = c.diff(); gain = delta.clip(lower=0); loss = (-delta).clip(lower=0)
|
||
for p in [7,14,21]:
|
||
ag=gain.rolling(p).mean(); al=loss.rolling(p).mean()
|
||
df[f'rsi_{p}'] = 100 - 100/(1+ag/al.replace(0,np.nan))
|
||
df['rsi_14_slope'] = df['rsi_14'].diff(5)
|
||
|
||
mid=c.rolling(20).mean(); std=c.rolling(20).std()
|
||
df['bb_pct'] = (c-(mid-2*std))/((mid+2*std)-(mid-2*std)).replace(0,np.nan)
|
||
df['bb_width'] = 4*std/mid
|
||
df['bb_width_chg'] = df['bb_width'].pct_change(10)
|
||
|
||
ema12=c.ewm(span=12,adjust=False).mean(); ema26=c.ewm(span=26,adjust=False).mean()
|
||
df['macd'] = (ema12-ema26)/c
|
||
df['macd_signal'] = df['macd'].ewm(span=9,adjust=False).mean()
|
||
df['macd_hist'] = df['macd']-df['macd_signal']
|
||
df['macd_hist_slope'] = df['macd_hist'].diff(3)
|
||
|
||
tr = pd.concat([h-l,(h-c.shift(1)).abs(),(l-c.shift(1)).abs()],axis=1).max(axis=1)
|
||
df['atr_pct'] = tr.rolling(14).mean()/c
|
||
df['atr_7'] = tr.rolling(7).mean()/c
|
||
df['atr_ratio'] = df['atr_7']/df['atr_pct'].replace(0,np.nan)
|
||
|
||
for p in [14,28]:
|
||
low_p=l.rolling(p).min(); high_p=h.rolling(p).max()
|
||
df[f'stoch_k_{p}'] = (c-low_p)/(high_p-low_p).replace(0,np.nan)*100
|
||
df['stoch_d_14'] = df['stoch_k_14'].rolling(3).mean()
|
||
|
||
for p in [1,3,5,10,20,60,120]:
|
||
df[f'ret_{p}'] = c.pct_change(p)
|
||
|
||
df['vol_5'] = c.pct_change().rolling(5).std()
|
||
df['vol_20'] = c.pct_change().rolling(20).std()
|
||
df['vol_60'] = c.pct_change().rolling(60).std()
|
||
df['vol_ratio'] = df['vol_5']/df['vol_20'].replace(0,np.nan)
|
||
df['vol_trend'] = df['vol_20'].pct_change(20)
|
||
|
||
body = (c-o).abs()
|
||
df['body_pct'] = body/c
|
||
df['upper_shadow'] = (h-pd.concat([o,c],axis=1).max(axis=1))/c
|
||
df['lower_shadow'] = (pd.concat([o,c],axis=1).min(axis=1)-l)/c
|
||
df['body_vs_range'] = body/(h-l).replace(0,np.nan)
|
||
df['range_pct'] = (h-l)/c
|
||
bullish = (c>o).astype(int)
|
||
df['streak'] = bullish.groupby((bullish!=bullish.shift()).cumsum()).cumcount()+1
|
||
df['streak'] = df['streak'] * bullish - df['streak'] * (1-bullish)
|
||
df['engulf_ratio'] = body/body.shift(1).replace(0,np.nan)
|
||
|
||
for p in [20,60]:
|
||
df[f'high_{p}'] = h.rolling(p).max()
|
||
df[f'low_{p}'] = l.rolling(p).min()
|
||
df[f'pos_{p}'] = (c-df[f'low_{p}'])/(df[f'high_{p}']-df[f'low_{p}']).replace(0,np.nan)
|
||
|
||
# 5分钟
|
||
c5=c.resample('5min').last(); h5=h.resample('5min').max()
|
||
l5=l.resample('5min').min(); o5=o.resample('5min').first()
|
||
e5_8=c5.ewm(span=8,adjust=False).mean(); e5_21=c5.ewm(span=21,adjust=False).mean()
|
||
df['ema5m_fs'] = ((e5_8-e5_21)/c5).reindex(df.index, method='ffill')
|
||
d5=c5.diff(); g5=d5.clip(lower=0).rolling(14).mean(); l5r=(-d5).clip(lower=0).rolling(14).mean()
|
||
df['rsi5m'] = (100-100/(1+g5/l5r.replace(0,np.nan))).reindex(df.index, method='ffill')
|
||
tr5=pd.concat([h5-l5,(h5-c5.shift(1)).abs(),(l5-c5.shift(1)).abs()],axis=1).max(axis=1)
|
||
df['atr5m'] = (tr5.rolling(14).mean()/c5).reindex(df.index, method='ffill')
|
||
for p in [1,5,20]:
|
||
df[f'ret5m_{p}'] = c5.pct_change(p).reindex(df.index, method='ffill')
|
||
|
||
# 15分钟
|
||
c15=c.resample('15min').last()
|
||
e15=c15.ewm(span=21,adjust=False).mean()
|
||
df['ema15m_trend'] = ((c15-e15)/c15).reindex(df.index, method='ffill')
|
||
df['ret15m_5'] = c15.pct_change(5).reindex(df.index, method='ffill')
|
||
|
||
df['hour'] = df.index.hour; df['minute'] = df.index.minute
|
||
df['hour_sin'] = np.sin(2*np.pi*df['hour']/24)
|
||
df['hour_cos'] = np.cos(2*np.pi*df['hour']/24)
|
||
df['weekday'] = df.index.weekday
|
||
|
||
return df
|
||
|
||
def get_fcols(df):
|
||
exclude = {'ts','open','high','low','close','label','month',
|
||
'ema_5','ema_8','ema_13','ema_21','ema_50','ema_120',
|
||
'high_20','low_20','high_60','low_60'}
|
||
return [c for c in df.columns if c not in exclude
|
||
and df[c].dtype in ('float64','float32','int64','int32')]
|
||
|
||
def train_predict(df, fcols, fb, thresh, train_m=3):
|
||
future_ret = df['close'].shift(-fb)/df['close'] - 1
|
||
df['label'] = 0
|
||
df.loc[future_ret > thresh, 'label'] = 1
|
||
df.loc[future_ret < -thresh, 'label'] = -1
|
||
|
||
df['month'] = df.index.to_period('M')
|
||
months = sorted(df['month'].unique())
|
||
|
||
pl = pd.Series(index=df.index, dtype=float); pl[:] = 0.0
|
||
ps = pd.Series(index=df.index, dtype=float); ps[:] = 0.0
|
||
|
||
params = {
|
||
'objective':'multiclass','num_class':3,'metric':'multi_logloss',
|
||
'learning_rate':0.03,'num_leaves':63,'max_depth':8,
|
||
'min_child_samples':80,'subsample':0.7,'colsample_bytree':0.7,
|
||
'reg_alpha':0.3,'reg_lambda':0.3,'verbose':-1,'n_jobs':-1,'seed':42
|
||
}
|
||
|
||
for i in range(train_m, len(months)):
|
||
tm = months[i]; ts_ = months[i-train_m]
|
||
tr_mask = (df['month']>=ts_)&(df['month']<tm)
|
||
te_mask = df['month']==tm
|
||
tr_df = df[tr_mask].dropna(subset=fcols+['label'])
|
||
te_df = df[te_mask].dropna(subset=fcols)
|
||
if len(tr_df)<1000 or len(te_df)<100: continue
|
||
dt_ = lgb.Dataset(tr_df[fcols].values, label=tr_df['label'].values+1)
|
||
model = lgb.train(params, dt_, num_boost_round=300)
|
||
proba = model.predict(te_df[fcols].values)
|
||
pl.loc[te_df.index] = proba[:,2]
|
||
ps.loc[te_df.index] = proba[:,0]
|
||
|
||
return pl, ps
|
||
|
||
def backtest(df, pl, ps, prob_th, sl, tp, mh, gap=0.03):
|
||
NOTIONAL=10000.0; FEE=NOTIONAL*0.0006*2; REB=FEE*0.9; NFEE=FEE-REB
|
||
pos=0; op=0.0; ot=None; trades=[]
|
||
|
||
for i in range(len(df)):
|
||
dt=df.index[i]; p=df['close'].iloc[i]; p_l=pl.iloc[i]; p_s=ps.iloc[i]
|
||
if pos!=0 and ot is not None:
|
||
pp=(p-op)/op if pos==1 else (op-p)/op
|
||
hsec=(dt-ot).total_seconds()
|
||
hard_sl=max(sl*1.5,0.006)
|
||
if -pp>=hard_sl: trades.append((pos,op,p,NOTIONAL*pp,hsec,'hsl',ot,dt));pos=0;continue
|
||
if hsec>=200:
|
||
if -pp>=sl: trades.append((pos,op,p,NOTIONAL*pp,hsec,'sl',ot,dt));pos=0;continue
|
||
if pp>=tp: trades.append((pos,op,p,NOTIONAL*pp,hsec,'tp',ot,dt));pos=0;continue
|
||
if hsec>=mh: trades.append((pos,op,p,NOTIONAL*pp,hsec,'to',ot,dt));pos=0;continue
|
||
if pos==1 and p_s>prob_th+0.08:
|
||
trades.append((pos,op,p,NOTIONAL*pp,hsec,'ai',ot,dt));pos=0
|
||
elif pos==-1 and p_l>prob_th+0.08:
|
||
trades.append((pos,op,p,NOTIONAL*pp,hsec,'ai',ot,dt));pos=0
|
||
if pos==0:
|
||
if p_l>prob_th and p_l>p_s+gap: pos=1;op=p;ot=dt
|
||
elif p_s>prob_th and p_s>p_l+gap: pos=-1;op=p;ot=dt
|
||
if pos!=0:
|
||
p=df['close'].iloc[-1];dt=df.index[-1]
|
||
pp=(p-op)/op if pos==1 else (op-p)/op
|
||
trades.append((pos,op,p,NOTIONAL*pp,(dt-ot).total_seconds(),'end',ot,dt))
|
||
return trades
|
||
|
||
def score(trades):
|
||
if not trades: return 0, 0, 0, 0, {}
|
||
NFEE=1.2; n=len(trades)
|
||
tpnl=sum(t[3] for t in trades); net=tpnl-NFEE*n
|
||
wins=len([t for t in trades if t[3]>0]); wr=wins/n*100
|
||
cum=0;peak=0;dd=0
|
||
monthly=defaultdict(lambda:{'n':0,'net':0,'w':0})
|
||
for t in trades:
|
||
cum+=t[3]-NFEE
|
||
if cum>peak:peak=cum
|
||
if peak-cum>dd:dd=peak-cum
|
||
k=t[7].strftime('%Y-%m')
|
||
monthly[k]['n']+=1;monthly[k]['net']+=t[3]-NFEE
|
||
if t[3]>0:monthly[k]['w']+=1
|
||
pm=len([m for m in monthly.values() if m['net']>0])
|
||
return net, wr, pm, dd, monthly
|
||
|
||
def main():
|
||
t0=_time.time()
|
||
print("="*70, flush=True)
|
||
print(" AI快速优化 | 100U x 100倍 | LightGBM + 63特征", flush=True)
|
||
print("="*70, flush=True)
|
||
|
||
df = load_data()
|
||
df = add_features(df)
|
||
fcols = get_fcols(df)
|
||
print(f" {len(df):,} bars, {len(fcols)} features\n", flush=True)
|
||
|
||
# 预训练不同前瞻/阈值的模型（最耗时的部分）
|
||
model_configs = [
|
||
(10, 0.003, 3, "10bar/0.3%/3m"),
|
||
(10, 0.003, 4, "10bar/0.3%/4m"),
|
||
(10, 0.004, 3, "10bar/0.4%/3m"),
|
||
(15, 0.004, 3, "15bar/0.4%/3m"),
|
||
(20, 0.005, 3, "20bar/0.5%/3m"),
|
||
]
|
||
|
||
predictions = {}
|
||
for fb, thresh, tm, lbl in model_configs:
|
||
print(f" Training: {lbl}...", flush=True)
|
||
dfc = df.copy()
|
||
pl, ps = train_predict(dfc, fcols, fb, thresh, tm)
|
||
predictions[lbl] = (pl, ps)
|
||
# 快速检查
|
||
t_ = backtest(dfc, pl, ps, 0.45, 0.005, 0.008, 1800)
|
||
n_, _, _, _, _ = score(t_)
|
||
print(f" quick check: {len(t_)} trades, net={n_:+.0f}", flush=True)
|
||
|
||
# 扫描回测参数
|
||
print(f"\n Scanning backtest params...\n", flush=True)
|
||
bt_configs = [
|
||
# prob_th, sl, tp, max_hold, gap
|
||
(0.42, 0.005, 0.008, 1800, 0.02),
|
||
(0.45, 0.005, 0.008, 1800, 0.03),
|
||
(0.48, 0.005, 0.008, 1800, 0.03),
|
||
(0.45, 0.005, 0.010, 2400, 0.03),
|
||
(0.48, 0.005, 0.010, 2400, 0.03),
|
||
(0.50, 0.005, 0.010, 2400, 0.03),
|
||
(0.45, 0.006, 0.012, 2400, 0.03),
|
||
(0.48, 0.006, 0.012, 3600, 0.03),
|
||
(0.50, 0.006, 0.015, 3600, 0.03),
|
||
(0.45, 0.004, 0.008, 1800, 0.03),
|
||
(0.42, 0.004, 0.006, 1200, 0.02),
|
||
]
|
||
|
||
results = []
|
||
for mlbl, (pl, ps) in predictions.items():
|
||
for prob_th, sl, tp, mh, gap in bt_configs:
|
||
trades = backtest(df, pl, ps, prob_th, sl, tp, mh, gap)
|
||
net, wr, pm, dd, monthly = score(trades)
|
||
n = len(trades)
|
||
if n > 0:
|
||
results.append({
|
||
'model': mlbl, 'prob': prob_th, 'sl': sl, 'tp': tp,
|
||
'mh': mh, 'gap': gap, 'n': n, 'net': net,
|
||
'wr': wr, 'pm': pm, 'dd': dd, 'monthly': monthly
|
||
})
|
||
|
||
# 按净利排序
|
||
results.sort(key=lambda x: x['net'], reverse=True)
|
||
|
||
# 打印Top 15
|
||
print(f"{'='*100}", flush=True)
|
||
print(f" TOP 15 配置 (100U保证金 x 100倍)", flush=True)
|
||
print(f"{'='*100}", flush=True)
|
||
print(f" {'#':>2} {'模型':<18} {'概率':>4} {'SL':>4} {'TP':>4} {'MH':>5} {'gap':>4} {'交易':>5} {'年净利':>8} {'月均':>6} {'胜率':>5} {'盈月':>4} {'回撤':>6}", flush=True)
|
||
print(f" {'-'*95}", flush=True)
|
||
|
||
for i, r in enumerate(results[:15]):
|
||
mavg = r['net']/12
|
||
print(f" {i+1:>2} {r['model']:<18} {r['prob']:.2f} {r['sl']*100:.1f}% {r['tp']*100:.1f}% {r['mh']:>5} {r['gap']:.2f} {r['n']:>5} {r['net']:>+8.0f} {mavg:>+6.0f} {r['wr']:>4.1f}% {r['pm']:>2}/12 {r['dd']:>6.0f}", flush=True)
|
||
|
||
# 最佳方案详情
|
||
best = results[0]
|
||
print(f"\n{'='*70}", flush=True)
|
||
print(f" 最佳方案: {best['model']}", flush=True)
|
||
print(f" 概率>{best['prob']} SL={best['sl']*100:.1f}% TP={best['tp']*100:.1f}% MH={best['mh']}s gap={best['gap']}", flush=True)
|
||
print(f" 年净利: {best['net']:+.0f} USDT = 月均 {best['net']/12:+.0f} USDT", flush=True)
|
||
print(f" 交易: {best['n']}笔 | 胜率: {best['wr']:.1f}% | 盈利月: {best['pm']}/12 | 回撤: {best['dd']:.0f}", flush=True)
|
||
print(f"{'='*70}", flush=True)
|
||
|
||
print(f"\n 月度明细:", flush=True)
|
||
for m in sorted(best['monthly'].keys()):
|
||
d = best['monthly'][m]
|
||
wr_m = d['w']/d['n']*100 if d['n']>0 else 0
|
||
s = "盈利" if d['net']>0 else "亏损"
|
||
print(f" {m}: {d['n']:>4}笔 {d['net']:>+8.0f}U 胜率{wr_m:.0f}% [{s}]", flush=True)
|
||
|
||
# 对比
|
||
print(f"\n 对比:", flush=True)
|
||
print(f" 纯EMA策略: +1196/年 = +100/月 (227笔)", flush=True)
|
||
print(f" AI v1基线: +4801/年 = +400/月 (923笔)", flush=True)
|
||
print(f" AI v2优化: {best['net']:+.0f}/年 = {best['net']/12:+.0f}/月 ({best['n']}笔)", flush=True)
|
||
if best['net'] > 4801:
|
||
print(f" v2 vs v1提升: {(best['net']/4801-1)*100:+.0f}%", flush=True)
|
||
|
||
elapsed = _time.time()-t0
|
||
print(f"\n 耗时: {elapsed:.0f}s", flush=True)
|
||
print(f"{'='*70}", flush=True)
|
||
|
||
if __name__=='__main__':
|
||
main()
|