1{
2  "op": "replace",
3  "filename": "qlib/custom_model.py",
4  "start_line": 16,
5  "end_line": 103,
6  "content": "# =====================================================================\n# EDITABLE: CustomModel — implement your stock prediction model here\n# =====================================================================\nimport lightgbm as lgb\n\n\nclass CustomModel(Model):\n    \"\"\"Drift-Aware Ensemble with Distribution-Calibrated Temporal Weighting.\n\n    Novel contributions:\n    1. Distribution-calibrated sample weighting based on quarterly feature-space\n       similarity to the validation period (targets concept drift directly)\n    2. Synergistic temporal-distributional weight product for ensemble diversity\n    3. Validation IC-weighted ensemble aggregation across drift-aware members\n    \"\"\"\n\n    def __init__(self):\n        super().__init__()\n        self.models, self.ensw, self.fitted = [], [], False\n        self.params = {\n            \"objective\": \"mse\", \"colsample_bytree\": 0.8879, \"learning_rate\": 0.2,\n            \"subsample\": 0.8789, \"lambda_l1\": 205.6999, \"lambda_l2\": 580.9768,\n            \"max_depth\": 8, \"num_leaves\": 210, \"num_threads\": 20, \"verbosity\": -1,\n        }\n\n    def _date_ic(self, preds, labels, index):\n        \"\"\"Compute mean daily Information Coefficient (Pearson).\"\"\"\n        df = pd.DataFrame({\"p\": preds, \"l\": labels}, index=index)\n        ics = [g[\"p\"].corr(g[\"l\"]) for _, g in df.groupby(level=0) if len(g) > 5]\n        return np.nanmean(ics) if ics else 0.0\n\n    def _dist_weights(self, x_np, dates, x_va_np):\n        \"\"\"Novel: per-sample weights from quarterly feature distribution drift.\n\n        Computes standardized mean distance between each training quarter's\n        feature distribution and the validation distribution, then converts\n        to exponential similarity scores as sample weights.\"\"\"\n        va_mu = np.nanmean(x_va_np, axis=0)\n        va_sd = np.nanstd(x_va_np, axis=0) + 1e-8\n        qtrs = dates.to_period('Q')\n        qd = {}\n        for q in qtrs.unique():\n            m = qtrs == q\n            qd[q] = np.mean(((np.nanmean(x_np[m], axis=0) - va_mu) / va_sd) ** 2)\n        d = np.array([qd[q] for q in qtrs])\n        return np.exp(-d / (np.median(d) + 1e-8))\n\n    def fit(self, dataset: DatasetH):\n        df_tr = dataset.prepare(\"train\", col_set=[\"feature\", \"label\"], data_key=DataHandlerLP.DK_L)\n        df_va = dataset.prepare(\"valid\", col_set=[\"feature\", \"label\"], data_key=DataHandlerLP.DK_L)\n        x_tr, y_tr = df_tr[\"feature\"], df_tr[\"label\"].values.ravel()\n        x_va, y_va = df_va[\"feature\"], df_va[\"label\"].values.ravel()\n        dates = x_tr.index.get_level_values(0)\n        ud = sorted(dates.unique()); nd = len(ud)\n        d2i = {d: i for i, d in enumerate(ud)}\n        si = np.array([d2i[d] for d in dates])\n        # Four weighting schemes: uniform, temporal, distributional, synergistic\n        w_u = np.ones(len(x_tr))\n        w_e = np.exp(-(nd - 1 - si) * np.log(2) / (nd * 0.35))\n        w_d = self._dist_weights(x_tr.values, dates, x_va.values)\n        w_c = w_e * w_d  # Novel: multiplicative temporal x distributional\n        self.models, ics = [], []\n        for w in [w_u, w_e, w_d, w_c]:\n            w = w / (w.mean() + 1e-8)\n            dt = lgb.Dataset(x_tr.values, label=y_tr, weight=w, free_raw_data=False)\n            dv = lgb.Dataset(x_va.values, label=y_va, free_raw_data=False)\n            m = lgb.train(self.params, dt, num_boost_round=1000,\n                         valid_sets=[dt, dv], valid_names=[\"train\", \"valid\"],\n                         callbacks=[lgb.early_stopping(50), lgb.log_evaluation(200)])\n            self.models.append(m)\n            ics.append(self._date_ic(m.predict(x_va.values), y_va, x_va.index))\n        ic = np.clip(np.array(ics), 0, None)\n        self.ensw = ic / (ic.sum() + 1e-8) if ic.sum() > 0 else np.ones(4) / 4\n        self.fitted = True\n\n    def predict(self, dataset: DatasetH, segment=\"test\"):\n        if not self.fitted:\n            raise ValueError(\"Not fitted!\")\n        x = dataset.prepare(segment, col_set=\"feature\", data_key=DataHandlerLP.DK_I)\n        p = sum(w * m.predict(x.values) for w, m in zip(self.ensw, self.models))\n        return pd.Series(p, index=x.index, name=\"score\")\n"
7}

1{
2  "op": "replace",
3  "filename": "qlib/custom_model.py",
4  "start_line": 16,
5  "end_line": 96,
6  "content": "# =====================================================================\n# EDITABLE: CustomModel — implement your stock prediction model here\n# =====================================================================\nimport lightgbm as lgb\n\n\nclass CustomModel(Model):\n    \"\"\"Drift-Aware Ensemble with Distribution-Calibrated Temporal Weighting.\n\n    Novel contributions:\n    1. Distribution-calibrated sample weighting via quarterly feature-space\n       similarity to validation (targets concept drift directly)\n    2. Synergistic temporal x distributional weight product\n    3. Recent-emphasis model for capturing late-period patterns\n    4. ICIR-weighted adaptive-temperature softmax ensemble aggregation\n    \"\"\"\n\n    def __init__(self):\n        super().__init__()\n        self.models, self.ensw, self.fitted = [], [], False\n        self.params = {\n            \"objective\": \"mse\", \"colsample_bytree\": 0.8879, \"learning_rate\": 0.2,\n            \"subsample\": 0.8789, \"lambda_l1\": 205.6999, \"lambda_l2\": 580.9768,\n            \"max_depth\": 8, \"num_leaves\": 210, \"num_threads\": 20, \"verbosity\": -1,\n        }\n\n    def _icir(self, preds, labels, index):\n        \"\"\"Compute ICIR (IC Information Ratio) for robust model ranking.\"\"\"\n        df = pd.DataFrame({\"p\": preds, \"l\": labels}, index=index)\n        ics = [g[\"p\"].corr(g[\"l\"]) for _, g in df.groupby(level=0) if len(g) > 5]\n        ics = [x for x in ics if not np.isnan(x)]\n        if len(ics) < 2:\n            return 0.0\n        return np.mean(ics) / (np.std(ics) + 1e-8)\n\n    def _dist_weights(self, x_np, dates, x_va_np):\n        \"\"\"Novel: per-sample weights from quarterly feature distribution drift.\n        Maps each training quarter to an exponential similarity score based on\n        standardized mean distance to the validation feature distribution.\"\"\"\n        va_mu = np.nanmean(x_va_np, axis=0)\n        va_sd = np.nanstd(x_va_np, axis=0) + 1e-8\n        qtrs = dates.to_period('Q')\n        qd = {}\n        for q in qtrs.unique():\n            m = qtrs == q\n            qd[q] = np.mean(((np.nanmean(x_np[m], axis=0) - va_mu) / va_sd) ** 2)\n        d = np.array([qd[q] for q in qtrs])\n        return np.exp(-d / (np.median(d) + 1e-8))\n\n    def fit(self, dataset: DatasetH):\n        df_tr = dataset.prepare(\"train\", col_set=[\"feature\", \"label\"], data_key=DataHandlerLP.DK_L)\n        df_va = dataset.prepare(\"valid\", col_set=[\"feature\", \"label\"], data_key=DataHandlerLP.DK_L)\n        x_tr, y_tr = df_tr[\"feature\"], df_tr[\"label\"].values.ravel()\n        x_va, y_va = df_va[\"feature\"], df_va[\"label\"].values.ravel()\n        dates = x_tr.index.get_level_values(0)\n        ud = sorted(dates.unique()); nd = len(ud)\n        d2i = {d: i for i, d in enumerate(ud)}\n        si = np.array([d2i[d] for d in dates])\n        # Five weighting schemes for temporal ensemble diversity\n        w_u = np.ones(len(x_tr))\n        w_e = np.exp(-(nd - 1 - si) * np.log(2) / (nd * 0.35))\n        w_d = self._dist_weights(x_tr.values, dates, x_va.values)\n        w_c = w_e * w_d  # Novel: synergistic temporal x distributional\n        w_r = np.where(si >= int(nd * 0.6), 1.0, 0.05)  # Recent 40% emphasis\n        self.models, scores = [], []\n        for w in [w_u, w_e, w_d, w_c, w_r]:\n            w = w / (w.mean() + 1e-8)\n            dt = lgb.Dataset(x_tr.values, label=y_tr, weight=w, free_raw_data=False)\n            dv = lgb.Dataset(x_va.values, label=y_va, free_raw_data=False)\n            m = lgb.train(self.params, dt, num_boost_round=1000,\n                         valid_sets=[dt, dv], valid_names=[\"train\", \"valid\"],\n                         callbacks=[lgb.early_stopping(50), lgb.log_evaluation(200)])\n            self.models.append(m)\n            scores.append(self._icir(m.predict(x_va.values), y_va, x_va.index))\n        # Adaptive-temperature softmax from ICIR scores\n        s = np.clip(np.array(scores), 0, None)\n        if s.sum() > 0:\n            tau = np.std(s) + 1e-8\n            self.ensw = np.exp(s / tau)\n            self.ensw /= self.ensw.sum()\n        else:\n            self.ensw = np.ones(5) / 5\n        self.fitted = True\n\n    def predict(self, dataset: DatasetH, segment=\"test\"):\n        if not self.fitted:\n            raise ValueError(\"Not fitted!\")\n        x = dataset.prepare(segment, col_set=\"feature\", data_key=DataHandlerLP.DK_I)\n        p = sum(w * m.predict(x.values) for w, m in zip(self.ensw, self.models))\n        return pd.Series(p, index=x.index, name=\"score\")\n"
7}

1{
2  "op": "replace",
3  "filename": "qlib/custom_model.py",
4  "start_line": 42,
5  "end_line": 49,
6  "content": "    def _date_ic(self, preds, labels, index):\n        \"\"\"Compute mean daily IC for ensemble weight calibration.\"\"\"\n        df = pd.DataFrame({\"p\": preds, \"l\": labels}, index=index)\n        ics = [g[\"p\"].corr(g[\"l\"]) for _, g in df.groupby(level=0) if len(g) > 5]\n        return np.nanmean(ics) if ics else 0.0"
7}

1{
2  "op": "replace",
3  "filename": "qlib/custom_model.py",
4  "start_line": 86,
5  "end_line": 94,
6  "content": "            scores.append(self._date_ic(m.predict(x_va.values), y_va, x_va.index))\n        # Proportional IC-weighted ensemble aggregation\n        ic = np.clip(np.array(scores), 0, None)\n        self.ensw = ic / (ic.sum() + 1e-8) if ic.sum() > 0 else np.ones(5) / 5"
7}