{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "33b0ae62",
   "metadata": {},
   "source": [
    "# TI 2026 in 5 charts — starter notebook\n",
    "\n",
    "538 pro Dota 2 games (every game of **TI 2026**, TI 2025, EWC 2025/2026), every 24-step\n",
    "Captains-Mode draft, plus a column nobody else publishes: **pre-series win\n",
    "probabilities archived before each TI 2026 series was played**.\n",
    "\n",
    "Dataset curated by [batru.gg](https://batru.gg) · data via the OpenDota API · CC BY 4.0."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ff8734c9",
   "metadata": {},
   "outputs": [],
   "source": [
    "import glob, os\n",
    "import pandas as pd\n",
    "\n",
    "# Works on Kaggle (any dataset slug), Hugging Face checkouts, or a local copy.\n",
    "hits = glob.glob(\"/kaggle/input/*/games.parquet\")\n",
    "BASE = os.path.dirname(hits[0]) if hits else \".\"\n",
    "\n",
    "games = pd.read_parquet(f\"{BASE}/games.parquet\")\n",
    "steps = pd.read_parquet(f\"{BASE}/draft_steps.parquet\")\n",
    "heroes = pd.read_csv(f\"{BASE}/hero_stats.csv\")\n",
    "games.groupby(\"tournament\").size()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3299f0b5",
   "metadata": {},
   "source": [
    "## 1. The 100% club — Treant Protector was picked or banned in ALL 148 TI 2026 games"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1ad2c53e",
   "metadata": {},
   "outputs": [],
   "source": [
    "ti = heroes[heroes.tournament == \"ti-2026\"].copy()\n",
    "n_games = games[games.tournament == \"ti-2026\"].match_id.nunique()\n",
    "top = ti.sort_values(\"contest_rate\", ascending=False).head(12)\n",
    "ax = top.set_index(\"hero\").contest_rate.mul(100).plot.barh(figsize=(8, 5), title=f\"Most contested heroes, TI 2026 ({n_games} games)\")\n",
    "ax.invert_yaxis(); ax.set_xlabel(\"% of games picked or banned\");"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "008c90e0",
   "metadata": {},
   "source": [
    "## 2. The popularity trap — most-picked ≠ best\n",
    "\n",
    "Hoodwink was the joint-most-picked hero of TI 2026 (73 games, tied with Mirana)\n",
    "and won only 41% of them. Meanwhile Earth Spirit was contested in 92.6% of\n",
    "drafts **and** delivered a ~66% win rate."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "86ed9e35",
   "metadata": {},
   "outputs": [],
   "source": [
    "t = ti[ti.picks >= 15].copy()\n",
    "t[\"win_rate\"] = t.wins_when_picked / t.picks\n",
    "t.sort_values(\"win_rate\", ascending=False)[[\"hero\", \"picks\", \"bans\", \"win_rate\"]].round(3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2231d046",
   "metadata": {},
   "outputs": [],
   "source": [
    "ax = t.plot.scatter(x=\"picks\", y=\"win_rate\", figsize=(8, 5), title=\"TI 2026: picks vs win rate (15+ picks)\")\n",
    "ax.axhline(0.5, ls=\"--\", lw=1)\n",
    "for _, r in t.iterrows():\n",
    "    ax.annotate(r.hero, (r.picks, r.win_rate), fontsize=7, xytext=(3, 3), textcoords=\"offset points\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "28314cf2",
   "metadata": {},
   "source": [
    "## 3. First pick and side — smaller edges than folklore says"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7455f895",
   "metadata": {},
   "outputs": [],
   "source": [
    "g = games.assign(\n",
    "    fp_won=lambda d: (d.first_pick_side == \"radiant\") == d.radiant_win,\n",
    ")\n",
    "pd.DataFrame({\n",
    "    \"first_pick_wr\": g.groupby(\"tournament\").fp_won.mean(),\n",
    "    \"radiant_wr\": g.groupby(\"tournament\").radiant_win.mean(),\n",
    "    \"games\": g.groupby(\"tournament\").size(),\n",
    "}).round(3)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c4bae9b9",
   "metadata": {},
   "source": [
    "## 4. Draft order — where in the draft does each hero go?\n",
    "\n",
    "`step` runs 1–24 in standard Captains-Mode order. Low mean step = ban-phase-1\n",
    "priority target; high mean step = late-pick flex / counter-pick material."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7277dab7",
   "metadata": {},
   "outputs": [],
   "source": [
    "ti_steps = steps[steps.tournament == \"ti-2026\"]\n",
    "order = ti_steps.groupby(\"hero\").agg(times=(\"step\", \"size\"), mean_step=(\"step\", \"mean\")).query(\"times >= 20\")\n",
    "order.sort_values(\"mean_step\").round(1).head(15)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1e1eb0d0",
   "metadata": {},
   "source": [
    "## 5. The honesty column — scoring the archived model predictions\n",
    "\n",
    "86 TI 2026 games carry a win probability **archived before the series started**\n",
    "(`model_pre_series_p_team1`). Because they were recorded pre-game, you can score\n",
    "them fairly — no hindsight. Brier score: 0.25 = always saying 50/50; lower is better.\n",
    "\n",
    "TI 2026 was an upset-heavy event (the champion came through the entire lower\n",
    "bracket), which is exactly why archived-not-backfilled predictions matter: the\n",
    "2025/EWC rows carry `null` instead of fabricated retro-predictions. The model's\n",
    "full public track record lives at [batru.gg/dota2/benchmark](https://batru.gg/dota2/benchmark).\n",
    "\n",
    "**Ideas to build on this dataset:** predict the winner from the draft alone\n",
    "(24 hero slots -> outcome); model hero priority as a sequence problem; detect\n",
    "meta shifts between tournaments/patches; find over/under-drafted heroes."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9903bad2",
   "metadata": {},
   "outputs": [],
   "source": [
    "m = games.dropna(subset=[\"model_pre_series_p_team1\"]).copy()\n",
    "m[\"team1_won\"] = (m.winner_team == m.series_team1).astype(float)\n",
    "brier = ((m.model_pre_series_p_team1 - m.team1_won) ** 2).mean()\n",
    "acc = ((m.model_pre_series_p_team1 > 0.5) == (m.team1_won == 1)).mean()\n",
    "print(f\"n={len(m)}  brier={brier:.4f} (0.25 = coin-flip baseline)  per-game acc={acc:.1%}\")"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}