src/streamlit_app.py

# =============================================================
# ⚽ XGBoost 3-Class 예측 + 유사 경기 분포 (Full Integration, Mobile-Fix)
# =============================================================
import streamlit as st
import pandas as pd
import numpy as np
import joblib

# ===============================
# 앱 기본 설정 + 모바일 표시 버그 대응 CSS/JS
# ===============================
st.set_page_config(page_title="⚽ XGBoost 예측 + 유사 경기 탐색기", layout="wide")

# 🔧 모바일에서 텍스트/메트릭이 잘리지 않도록 overflow/zoom/글꼴 고정 + 초기 리사이즈 트리거
st.markdown(
    """
    <style>
    .stApp, .block-container, [data-testid="stVerticalBlock"], [data-testid="column"] {
        overflow: visible !important;
    }
    @media (max-width: 768px) {
        .block-container {
            padding-left: 1rem !important;
            padding-right: 1rem !important;
        }
        .prob-pill {
            font-size: 1.1rem;
            color: #ffffff;
            background: #262730;
            padding: 6px 10px;
            border-radius: 8px;
            display: inline-block;
            margin: 4px 6px 4px 0;
            line-height: 1.2;
        }
        .prob-section h3 {
            margin-top: 0.4rem;
            margin-bottom: 0.4rem;
        }
    }
    </style>
    <script>
    window.addEventListener('load', () => {
      setTimeout(() => { window.dispatchEvent(new Event('resize')); }, 150);
    });
    </script>
    """,
    unsafe_allow_html=True,
)

st.title("⚽ XGBoost 예측 + 유사 경기 분포 + 유사 경기 탐색기")

EQ_DECIMALS = 2
def eq(a, b, decimals=EQ_DECIMALS):
    return np.round(a, decimals) == np.round(b, decimals)

# ===============================
# Feature 목록 (65피처 통합)
# ===============================
expected_cols_65 = [
    'norm_win','norm_draw','norm_lose','mean_odds','std_odds','cv_odds',
    'p_win','p_draw','p_lose','overround','entropy','spread','spread_draw',
    'odds_ratio_wd','odds_ratio_wl','odds_ratio_dl','draw_prob_ratio','draw_ratio',
    'draw_prob_gap','fav_gap','fav_draw_gap','fav_diff','draw_gap_mean',
    'rank_win','rank_draw','rank_lose','p_win_norm','p_draw_norm','p_lose_norm',
    'ev_win','ev_draw','ev_lose','draw_vs_avg','draw_vs_max','cv_spread','cv_draw_gap',
    'draw_margin','fav_ratio','draw_skew','log_spread','draw_entropy_component','dominance_score',
    'hmean_odds','hstd_odds','hcv_odds','hentropy','hspread','hspread_draw',
    'hp_win','hp_draw','hp_lose','hp_win_norm','hp_draw_norm','hp_lose_norm','hoverround',
    'diff_win_prob','diff_draw_prob','diff_lose_prob','diff_draw_odds',
    'base_win_odds','base_draw_odds','base_lose_odds',
    'base_overround_ex','base_entropy_ex','base_spread_ex'
]

# ===============================
# Feature 생성
# ===============================
def build_feature_dict(win, draw, lose, hwin, hdraw, hlose):
    d = {}
    denom = (win+draw+lose)
    d['norm_win'], d['norm_draw'], d['norm_lose'] = win/denom, draw/denom, lose/denom
    d['mean_odds'] = np.mean([win,draw,lose])
    d['std_odds'] = np.std([win,draw,lose])
    d['cv_odds'] = d['std_odds']/d['mean_odds'] if d['mean_odds']>0 else 0
    d['p_win'], d['p_draw'], d['p_lose'] = 1/win, 1/draw, 1/lose
    p_tot = d['p_win'] + d['p_draw'] + d['p_lose']
    d['p_win_norm'], d['p_draw_norm'], d['p_lose_norm'] = d['p_win']/p_tot, d['p_draw']/p_tot, d['p_lose']/p_tot
    d['overround'] = p_tot
    d['entropy'] = -sum(x*np.log(x) for x in [d['p_win_norm'],d['p_draw_norm'],d['p_lose_norm']])
    d['spread'] = max(win,draw,lose)-min(win,draw,lose)
    d['spread_draw'] = abs(draw-(win+lose)/2)
    d['odds_ratio_wd'],d['odds_ratio_wl'],d['odds_ratio_dl']=win/draw,win/lose,draw/lose
    d['draw_prob_ratio']=d['p_draw']/max(d['p_win'],d['p_lose'])
    d['draw_ratio']=draw/min(win,lose)
    d['draw_prob_gap']=abs(d['p_draw']-(d['p_win']+d['p_lose'])/2)
    d['fav_gap']=abs(win-lose)
    d['fav_draw_gap']=abs(draw-min(win,lose))
    d['fav_diff']=abs(win-lose)
    d['draw_gap_mean']=abs(draw-d['mean_odds'])
    d['rank_win'],d['rank_draw'],d['rank_lose']=pd.Series([win,draw,lose]).rank().tolist()
    d['ev_win'],d['ev_draw'],d['ev_lose']=win*d['p_win_norm'],draw*d['p_draw_norm'],lose*d['p_lose_norm']
    d['draw_vs_avg']=draw/d['mean_odds']
    d['draw_vs_max']=draw/max(win,draw,lose)
    d['cv_spread']=d['spread']/d['mean_odds']
    d['cv_draw_gap']=d['fav_draw_gap']/d['mean_odds']
    d['draw_margin']=abs(draw-(win+lose)/2)
    d['fav_ratio']=min(win,lose)/max(win,lose)
    d['draw_skew']=(draw-win)-(lose-draw)
    d['log_spread']=np.log(max(win,draw,lose))-np.log(min(win,draw,lose))
    d['draw_entropy_component']=-d['p_draw_norm']*np.log(d['p_draw_norm'])
    d['dominance_score']=max(d['p_win_norm'],d['p_lose_norm'])-d['p_draw_norm']
    d['hmean_odds']=np.mean([hwin,hdraw,hlose])
    d['hstd_odds']=np.std([hwin,hdraw,hlose])
    d['hcv_odds']=d['hstd_odds']/d['hmean_odds'] if d['hmean_odds']>0 else 0
    p_h=1/np.array([hwin,hdraw,hlose]); p_hn=p_h/p_h.sum()
    d['hp_win'],d['hp_draw'],d['hp_lose']=p_h
    d['hp_win_norm'],d['hp_draw_norm'],d['hp_lose_norm']=p_hn
    d['hoverround']=p_h.sum()
    d['hentropy']=-np.sum(p_hn*np.log(p_hn))
    d['hspread']=max(hwin,hdraw,hlose)-min(hwin,hdraw,hlose)
    d['hspread_draw']=abs(hdraw-(hwin+hlose)/2)
    d['diff_win_prob']=d['p_win_norm']-d['hp_win_norm']
    d['diff_draw_prob']=d['p_draw_norm']-d['hp_draw_norm']
    d['diff_lose_prob']=d['p_lose_norm']-d['hp_lose_norm']
    d['diff_draw_odds']=hdraw-draw
    d['base_win_odds'],d['base_draw_odds'],d['base_lose_odds']=win,draw,lose
    d['base_overround_ex'],d['base_entropy_ex'],d['base_spread_ex']=p_tot,d['entropy'],d['spread']
    return d

def build_feature_frame(win, draw, lose, hwin, hdraw, hlose):
    d = build_feature_dict(win, draw, lose, hwin, hdraw, hlose)
    df = pd.DataFrame([d])
    return df[expected_cols_65]

# ===============================
# 모델 로드 (XGBoost)
# ===============================
@st.cache_resource
def load_models():
    base = joblib.load("xgb_model_wdl_softmax.pkl")
    hand = joblib.load("xgb_model_handicap_65f.pkl")
    enc  = joblib.load("label_encoder_handicap.pkl")
    return base, hand, enc

model_base, model_hand, encoder_hand = load_models()

# ===============================
# 예측
# ===============================
def predict_all(win, draw, lose, hwin, hdraw, hlose):
    df_feat = build_feature_frame(win, draw, lose, hwin, hdraw, hlose)
    probs_base = model_base.predict_proba(df_feat.values)[0]
    probs_hand = model_hand.predict_proba(df_feat.values)[0]
    return dict(zip(["승","무","패"], probs_base)), dict(zip(["핸디 승","핸디 무","핸디 패"], probs_hand))

# ===============================
# DB 로드
# ===============================
@st.cache_data
def load_db():
    df = pd.read_excel("proto_core_65_fastsearch.xlsx", engine="openpyxl")
    for c in ["승","무","패","핸디 승","핸디 무","핸디 패"]:
        df[c] = pd.to_numeric(df[c], errors="coerce")
    return df
df = load_db()

# ===============================
# 사이드바: 입력 배당
# ===============================
st.sidebar.header("⚙️ 입력 배당")
odds_str = st.sidebar.text_input("배당 (승/무/패/핸승/핸무/핸패)", "2.05/3.35/3.45/3.65/3.75/1.90")

try:
    win, draw, lose, hwin, hdraw, hlose = map(float, odds_str.split("/"))
except:
    st.error("형식 오류! 예: 2.05/3.35/3.45/3.65/3.75/1.90")
    st.stop()

# 먼저 유사 경기 탐색기 섹션을 렌더 (모바일 렌더 순서 안정화)
st.header("🔍 유사 경기 탐색기 (정배당 일치 포함)")
base_odds, hand_odds = [win, draw, lose], [hwin, hdraw, hlose]
base_min_idx_in = np.argmin(base_odds)
hand_min_idx_in = np.argmin(hand_odds)
base_min_val, hand_min_val = min(base_odds), min(hand_odds)
base_rev_in = base_odds[2] if base_min_idx_in == 0 else base_odds[0]
hand_rev_in = hand_odds[2] if hand_min_idx_in == 0 else hand_odds[0]
base_rank_in = pd.Series(base_odds).rank().tolist()
hand_rank_in = pd.Series(hand_odds).rank().tolist()

dfc = df.copy()
for c in ["승","무","패","핸디 승","핸디 무","핸디 패"]:
    dfc[c] = pd.to_numeric(dfc[c], errors="coerce")

dfc["base_min_idx"] = dfc[["승","무","패"]].values.argmin(axis=1)
dfc["hand_min_idx"] = dfc[["핸디 승","핸디 무","핸디 패"]].values.argmin(axis=1)
dfc["base_rev_val"] = np.where(dfc["base_min_idx"]==0, dfc["패"], dfc["승"])
dfc["hand_rev_val"] = np.where(dfc["hand_min_idx"]==0, dfc["핸디 패"], dfc["핸디 승"])
dfc["draw_rank"] = pd.DataFrame(dfc[["승","무","패"]]).rank(axis=1).iloc[:,1]
dfc["hand_draw_rank"] = pd.DataFrame(dfc[["핸디 승","핸디 무","핸디 패"]]).rank(axis=1).iloc[:,1]


st.markdown("### 🔗 복합 조건 선택")
use_combo1 = st.checkbox("기본 정배당 + 무 + 기본 역배당")
use_combo2 = st.checkbox("핸디 정배당 + 핸무 + 핸디 역배당")
use_combo3 = st.checkbox("기본 역배당 + 무")
use_combo4 = st.checkbox("핸디 역배당 + 핸무")
use_combo5 = st.checkbox("무 + 핸무")
use_combo6 = st.checkbox("무 + 핸무 + 핸디 정배 소수 첫째자리 일치")
use_combo7 = st.checkbox("무 + 핸무 + 기본 정배 소수 첫째자리 일치")

mask = (dfc["base_min_idx"]==base_min_idx_in)&(dfc["hand_min_idx"]==hand_min_idx_in)

draw_rank_in = base_rank_in[1]
hand_draw_rank_in = hand_rank_in[1]
mask &= (dfc["draw_rank"] == draw_rank_in)
mask &= (dfc["hand_draw_rank"] == hand_draw_rank_in)

if use_combo1:
    mask &= (eq(dfc["무"], draw)&eq(dfc["base_rev_val"], base_rev_in)&
             (np.round(dfc[["승","무","패"]].min(axis=1),2)==round(base_min_val,2)))

if use_combo2:
    mask &= (eq(dfc["핸디 무"], hdraw)&eq(dfc["hand_rev_val"], hand_rev_in)&
             (np.round(dfc[["핸디 승","핸디 무","핸디 패"]].min(axis=1),2)==round(hand_min_val,2)))

if use_combo3:
    mask &= eq(dfc["무"], draw)&eq(dfc["base_rev_val"], base_rev_in)

if use_combo4:
    mask &= eq(dfc["핸디 무"], hdraw)&eq(dfc["hand_rev_val"], hand_rev_in)

if use_combo5:
    mask &= eq(dfc["무"], draw)&eq(dfc["핸디 무"], hdraw)

if use_combo6:
    mask &= (eq(dfc["무"], draw)&eq(dfc["핸디 무"], hdraw)&
             (np.floor(dfc[["핸디 승","핸디 무","핸디 패"]].min(axis=1)*10)/10.0==np.floor(hand_min_val*10)/10.0))

if use_combo7:
    mask &= (
        eq(dfc["무"], draw) & eq(dfc["핸디 무"], hdraw) &
        (np.floor(dfc[["승","무","패"]].min(axis=1)*10)/10.0 == np.floor(base_min_val*10)/10.0)
    )    

df_sim = dfc.loc[mask].copy().reset_index(drop=True)
st.subheader("✅ 유사 경기 결과")

if df_sim.empty:
    st.warning("❌ 조건을 만족하는 유사 경기가 없습니다.")
else:
    cols_pref = ["일자","리그","홈팀","원정팀","승","무","패","핸디 승","핸디 무","핸디 패","결과","핸디결과"]
    cols_show = [c for c in cols_pref if c in df_sim.columns]
    if "일자" in df_sim.columns:
        df_sim["일자"] = pd.to_datetime(df_sim["일자"], errors="coerce").dt.strftime("%Y-%m-%d")
    st.dataframe(df_sim[cols_show], use_container_width=True)

    if "결과" in df_sim.columns or "핸디결과" in df_sim.columns:
        st.markdown("### 📊 결과 분포")
    if "결과" in df_sim.columns:
        st.write("**기본 시장 결과 분포:**", df_sim["결과"].value_counts().to_dict())
    if "핸디결과" in df_sim.columns:
        st.write("**핸디캡 시장 결과 분포:**", df_sim["핸디결과"].value_counts().to_dict())

st.divider()


# ===============================
# 🔢 확률 표시 (모바일 완전 호환)
# ===============================
base_probs, hand_probs = predict_all(win, draw, lose, hwin, hdraw, hlose)

mobile_text_mode = st.toggle("📱 모바일 강제 호환 모드(텍스트만)", value=True)

st.markdown('<div class="prob-section">', unsafe_allow_html=True)
st.markdown("### ⚽ 기본 승/무/패 확률")
if mobile_text_mode:
    for k, emoji in zip(["승","무","패"], ["🟢","🟡","🔴"]):
        st.markdown(f"{emoji} **{k}** : {base_probs[k]*100:.2f}%")
else:
    html = "".join([f'<span class="prob-pill"><b>{k}</b>: {base_probs[k]*100:.2f}%</span>'
                    for k in ["승","무","패"]])
    st.markdown(html, unsafe_allow_html=True)

st.markdown("### 🎯 핸디캡 승/무/패 확률")
if mobile_text_mode:
    for k, emoji in zip(["핸디 승","핸디 무","핸디 패"], ["🟢","🟡","🔴"]):
        st.markdown(f"{emoji} **{k}** : {hand_probs[k]*100:.2f}%")
else:
    html = "".join([f'<span class="prob-pill"><b>{k}</b>: {hand_probs[k]*100:.2f}%</span>'
                    for k in ["핸디 승","핸디 무","핸디 패"]])
    st.markdown(html, unsafe_allow_html=True)
st.markdown('</div>', unsafe_allow_html=True)

st.caption("ⓒ XGBoost 3-Class Softmax Models | 기본·핸디 65피처 통합형")