"""How much Wikipedia attention does a new album buy — and how long does it last? Run: python veri_hazirla.py (once; builds veri/*.csv) python analiz.py (prints the numbers, writes sonuclar.json, builds the dashboard) Question When an artist releases a studio album, how much do views of their English Wikipedia article rise, how fast does the rise fade, and is that more than an ordinary week? Design * Artists: 51 well-known artists, each matched to their English Wikipedia article and MusicBrainz id (see veri_hazirla.py). * Events: official studio albums (MusicBrainz release groups, primary type Album, no secondary type) with a full release date, Aug 2015 – Jul 2026, by a living artist, not overlapping another album by the same artist (veri_hazirla.py has the rules). * Window: day -35 .. +55 around the release date (day 0). baseline = median daily views on days -35 .. -8 (four weeks, ending a week before) peak = highest day in 0 .. 13, divided by baseline excess = sum over days 0 .. 55 of (views - baseline), divided by baseline = "extra baseline-days of attention in the first eight weeks" * Comparison: the same three numbers on ordinary Fridays (same artists, no album within 90 days, every 12 weeks) -> what a random week looks like. * Forecast check (rolling origin, chronological): predict an album's excess using only albums released before it. Simplest first: all_median median excess of every earlier album prev_album the same artist's previous album artist_median median of the same artist's earlier albums metric: mean absolute error of log10(excess) -> "typically off by a factor of x". """ from __future__ import annotations import json import re import sys from datetime import date, timedelta from pathlib import Path import numpy as np import pandas as pd HERE = Path(__file__).parent VERI = HERE / "veri" D0 = date(2015, 7, 1) OFF = list(range(-35, 56)) # window offsets BASE = (-35, -8) # inclusive PEAK = (0, 13) EXC = (0, 55) MIN_TRAIN = 20 # earlier albums needed before the first forecast def display(title: str) -> str: """Article title as a name: 'Drake_(musician)' -> 'Drake'. (MusicBrainz names differ: 'Ye', 'JAŸ-Z'.)""" return re.sub(r"\s*\(.*\)$", "", title.replace("_", " ")) def ds(i: int) -> str: return str(D0 + timedelta(days=int(i))) def load(): art = pd.read_csv(VERI / "sanatcilar.csv") alb = pd.read_csv(VERI / "albumler.csv") win = pd.read_csv(VERI / "pencereler.csv") pla = pd.read_csv(VERI / "plasebo.csv") mon = pd.read_csv(VERI / "aylik.csv") return art, alb, win, pla, mon def metrics(w: np.ndarray) -> tuple[float, float, float]: """w: 91 daily values for offsets -35..55.""" i = lambda o: o + 35 b = float(np.median(w[i(BASE[0]):i(BASE[1]) + 1])) peak = float(w[i(PEAK[0]):i(PEAK[1]) + 1].max()) / b exc = float((w[i(EXC[0]):i(EXC[1]) + 1] - b).sum()) / b return b, peak, exc def events(win: pd.DataFrame, art: pd.DataFrame) -> pd.DataFrame: cols = [f"d{o}" for o in OFF] m = np.array([metrics(r) for r in win[cols].to_numpy(float)]) e = win[["title", "date", "album"]].copy() e["baseline"], e["peak"], e["excess"] = m[:, 0], m[:, 1], m[:, 2] e["name"] = e["title"].map(display) return e.sort_values("date").reset_index(drop=True) def curve(win: pd.DataFrame, e: pd.DataFrame) -> dict: cols = [f"d{o}" for o in OFF] lift = win[cols].to_numpy(float) / e.set_index(["title", "date"]).loc[ list(zip(win.title, win.date)), "baseline"].to_numpy()[:, None] q = np.nanpercentile(lift, [25, 50, 75], axis=0) med = q[1] pk = int(np.argmax(med)) # index of the median peak half = 1 + (med[pk] - 1) / 2 after = np.where(med[pk:] <= half)[0] return {"q25": q[0].round(3).tolist(), "q50": med.round(3).tolist(), "q75": q[2].round(3).tolist(), "peak_day": OFF[pk], "peak_lift": round(float(med[pk]), 2), "half_life_days": int(after[0]) if len(after) else None, "lift_day7": round(float(med[OFF.index(7)]), 2), "lift_day28": round(float(med[OFF.index(28)]), 2), "lift_day55": round(float(med[OFF.index(55)]), 2), "pre_week_lift": round(float(np.median(med[OFF.index(-7):OFF.index(-1) + 1])), 2)} def backtest(e: pd.DataFrame) -> pd.DataFrame: rows = [] le = np.log10(e["excess"].clip(lower=0.5)) # excess can be ~0 or negative for flops for i in range(len(e)): past = e.iloc[:i] mine = past[past.title == e.title.iloc[i]] if i < MIN_TRAIN or mine.empty: continue lp, lm = le.iloc[:i], le[mine.index] rows.append({"title": e.title.iloc[i], "date": e.date.iloc[i], "album": e.album.iloc[i], "actual": le.iloc[i], "all_median": float(lp.median()), "prev_album": float(lm.iloc[-1]), "artist_median": float(lm.median())}) return pd.DataFrame(rows) def first_week_share(win: pd.DataFrame, e: pd.DataFrame) -> float: """Median share of an album's 8-week extra attention that arrives in days 0..6.""" b = e.set_index(["title", "date"]).loc[list(zip(win.title, win.date))] first = win[[f"d{o}" for o in range(0, 7)]].sum(axis=1).to_numpy() - 7 * b["baseline"].to_numpy() total = b["excess"].to_numpy() * b["baseline"].to_numpy() ok = total > 0 return round(float(np.median(first[ok] / total[ok]) * 100)) def spearman(a, b) -> float: return float(pd.Series(a).rank().corr(pd.Series(b).rank())) def main() -> dict: art, alb, win, pla, mon = load() e = events(win, art) cv = curve(win, e) bt = backtest(e) methods = ["all_median", "prev_album", "artist_median"] mae = {m: round(float((bt[m] - bt["actual"]).abs().mean()), 3) for m in methods} best = min(mae, key=mae.get) plc = pla[pla.peak > 0] peak_med, pla_med = float(e.peak.median()), float(plc.peak.median()) status = alb["status"].value_counts().to_dict() top = e.sort_values("excess", ascending=False).head(5) low = e.sort_values("excess").head(5) res = { "coverage": {"artists": int(art.shape[0]), "days": 4114, "first_day": ds(0), "last_day": ds(4113)}, "albums": {"candidates": int(len(alb)), "kept": int(status.get("kept", 0)), "dropped": {k: int(v) for k, v in status.items() if k != "kept"}, "artists_with_album": int(e.title.nunique())}, "former_titles": {"Drake": ["Drake (rapper)", "Drake (entertainer)"], "Jay-Z": ["Jay Z"], "BTS": ["Bangtan Boys", "BTS (band)"], "Blackpink": ["Black Pink"], "Wizkid": ["Wizkid (musician)"], "Sia": ["Sia (musician)", "Sia Furler"]}, "placebo_n": int(len(pla)), "placebo_artists": int(pla.title.nunique()), "peak": {"album_median": round(peak_med, 2), "album_q1": round(float(e.peak.quantile(.25)), 2), "album_q3": round(float(e.peak.quantile(.75)), 2), "placebo_median": round(pla_med, 2), "placebo_q90": round(float(plc.peak.quantile(.9)), 2), "share_placebo_above_album_median": round(float((plc.peak >= peak_med).mean() * 100), 1), "share_albums_above_placebo_q90": round(float((e.peak > plc.peak.quantile(.9)).mean() * 100), 1), "albums_below_1_5x": int((e.peak < 1.5).sum())}, "excess": {"album_median": round(float(e.excess.median()), 1), "album_q1": round(float(e.excess.quantile(.25)), 1), "album_q3": round(float(e.excess.quantile(.75)), 1), "placebo_median": round(float(pla.excess.median()), 1), "first_week_share_pct": first_week_share(win, e), "albums_negative": int((e.excess < 0).sum())}, "curve": cv, "size_vs_lift": {"spearman_baseline_peak": round(spearman(e.baseline, e.peak), 2), "spearman_baseline_excess": round(spearman(e.baseline, e.excess), 2)}, "backtest": {"n": int(len(bt)), "first": str(bt.date.min()), "last": str(bt.date.max()), "mae_log10": mae, "factor": {m: round(10 ** v, 2) for m, v in mae.items()}, "best": best, "improvement_pct": round(100 * (1 - mae[best] / mae["all_median"]))}, "top_excess": top[["name", "album", "date", "peak", "excess"]].round(1).to_dict("records"), "low_excess": low[["name", "album", "date", "peak", "excess"]].round(1).to_dict("records"), "events": e.round(3).to_dict("records"), "backtest_rows": bt.round(3).to_dict("records"), } (HERE / "sonuclar.json").write_text(json.dumps(res, indent=2, ensure_ascii=False, default=float)) if "--no-dashboard" not in sys.argv: from dashboard import build for p in build(res): print("written:", p) return res if __name__ == "__main__": out = main() for k in ["coverage", "albums", "placebo_n", "peak", "excess", "curve", "size_vs_lift", "backtest"]: v = out[k] if k == "curve": v = {x: y for x, y in v.items() if not x.startswith("q")} print(f"{k}: {v}") print("top:", [(r["name"], r["album"], r["excess"]) for r in out["top_excess"]])