"""DC2.1 feature build: per triggered coin, per checkpoint (trigger, +15m, +1h, +4h).""" import pandas as pd, numpy as np, bisect, sys DELAY = float(sys.argv[1]) if len(sys.argv) > 1 else 0.0 OUTF = sys.argv[2] if len(sys.argv) > 2 else "feat.csv.gz" R = "/tmp/claude-0/-home-claude/5101628c-bcc2-5fb2-9e13-0e431183bef0/scratchpad/" DISP = {"0x84157d3d24630c297849c84d49ae7f1805c1b60f", "0x03802649e469ed5680c58dff9e65425e22d21cc5"} L = pd.read_csv(R + "data/launches_15sep_4oct.csv.gz") X7 = pd.read_csv(R + "data/crossers500k_7d.csv.gz").set_index("tok").pk7 s = pd.read_csv(R + "raw/sw.csv.gz") coin = pd.read_csv(R + "raw/coin.csv.gz").set_index("token") bt = pd.read_csv(R + "raw/btime.csv.gz") ff = pd.read_csv(R + "raw/ff.csv.gz") cw = pd.read_csv(R + "raw/cwallet.csv.gz") cl = pd.read_csv(R + "raw/claunch.csv.gz") KT, KS, KMAX = bt.t.values, bt.spb.values, bt.k.max() def btt(b): b = np.asarray(b, dtype=np.int64); k = np.minimum(b // 1000, KMAX) return KT[k] + (b - k * 1000) * KS[k] for d in (s, coin): pass s["wallet"] = s.wallet.str.lower(); s["t"] = btt(s.b.values) coin["pull_t"] = np.where(coin.pull_b.notna(), btt(coin.pull_b.fillna(0).astype(np.int64).values), np.nan) DATA_END = s.t.max() L = L.set_index("tok") # ethusd per day: launch-file buy USD in first hour / wash buy ETH in first hour s0 = s.merge(L[["t0"]], left_on="token", right_index=True) h1 = s0[(s0.side == 1) & (s0.t <= s0.t0 + 3600)].groupby("token").eth.sum() r = (L.loc[h1.index, "bu_3600"] / h1).replace([np.inf, -np.inf], np.nan) r = r[(h1 > 0.05)] EU = r.groupby(L.loc[r.index, "day"]).median() print("ethusd per day", EU.round(0).to_dict(), flush=True) # wallet funding fund = {} for a, fb, via in zip(ff.addr.str.lower(), ff.fund_b, ff.via): fund[a] = (fb, str(via).lower()) for a, fb, via in zip(cw.addr.str.lower(), cw.fund_b, cw.via): if a not in fund or pd.isna(fund[a][0]): fund[a] = (fb, str(via).lower()) cwkind = dict(zip(cw.addr.str.lower(), cw.kind)) # creator history cl["creator"] = cl.creator.str.lower(); cl["t"] = btt(cl.b.values) cl_by_cr = cl.groupby("creator").t.apply(lambda x: np.sort(x.values)).to_dict() root_of = cl.drop_duplicates("creator").set_index("creator").root.str.lower().to_dict() L["cr"] = L.cr.str.lower() L["root"] = L.cr.map(root_of) L["t500"] = L.t0 + L.ts_500000; L["t1m"] = L.t0 + L["ts_1e+06"] past500_cr = L[L.t500.notna()].groupby("cr").t500.apply(lambda x: np.sort(x.values)).to_dict() past500_root = L[L.t500.notna() & L.root.notna()].groupby("root").t500.apply(lambda x: np.sort(x.values)).to_dict() # crew label of earlier $1M coins coin["t1m"] = L.t1m.reindex(coin.index) crew1m = coin[coin.t1m.notna()].groupby("crew").t1m.apply(lambda x: np.sort(x.values)).to_dict() # alumni: early (first 30 min) buyers of $1M coins, with that coin's $1M time g = s.groupby("token") first_t = g.t.min() s1 = s.merge(first_t.rename("ft"), left_on="token", right_index=True) ab = s1[(s1.side == 1) & (s1.t <= s1.ft + 1800)][["token", "wallet"]].drop_duplicates() ab["t1m"] = ab.token.map(L.t1m) ab = ab[ab.t1m.notna()] alum = ab.groupby("wallet").t1m.min().to_dict() # wallet -> earliest time it became alumni print("alumni wallets", len(alum), flush=True) def nb(a, x): return bisect.bisect_right(a, x) if a is not None else 0 CP = [("trig", 0), ("l30", 30), ("m15", 900), ("h1", 3600), ("h4", 14400)] rows = [] for tok, d in g: if tok not in L.index: continue lr = L.loc[tok]; day = lr.day; eu = EU.get(day, np.nan) d = d.sort_values(["b"]); t = d.t.values; px = d.px.values; side = d.side.values; eth = d.eth.values; w = d.wallet.values # usd-cap per unit px: median of mc_X / px at checkpoint ks = [] for X in (60, 300, 900, 1800, 3600): mc = lr.get("mc_%d" % X); i = np.searchsorted(t, lr.t0 + X, side="right") - 1 if pd.notna(mc) and i >= 0 and px[i] > 0: ks.append(mc / px[i]) if not ks or pd.isna(lr.ts_250000): continue K = np.median(ks) # unique buyers cumulative seen = set(); ub = np.zeros(len(t), int) for i in range(len(t)): if side[i] == 1: seen.add(w[i]) ub[i] = len(seen) i100 = np.argmax(ub >= 100) if ub[-1] >= 100 else None if i100 is None: continue t_trig = max(lr.t0 + lr.ts_250000, t[i100]) itr = np.searchsorted(t, t_trig, side="left") if itr >= len(t): continue t_trig = t[itr] pull_t = coin.pull_t.get(tok, np.nan); crew = coin.crew.get(tok, None) usd = eth * eu sellable = (side == -1) & (usd >= 100) cr = lr.cr; root = lr.root cr_fund = fund.get(cr, (np.nan, "")) for cpn, off in CP: tc = t_trig + off if pd.notna(pull_t) and pull_t <= tc: rows.append(dict(tok=tok, cp=cpn, alive=0, day=day)); continue if tc > DATA_END: continue ic = np.searchsorted(t, tc, side="right") - 1 pxc = px[ic]; capc = K * pxc; cap_eth = capc / eu m = slice(0, ic + 1) T, P, S, E, W = t[m], px[m], side[m], eth[m], w[m] last1h = T > tc - 3600; last15 = T > tc - 900 # holdings tokn = S * E / P hold = pd.Series(tokn).groupby(W).sum() pos = hold[hold > 0] hold1 = pd.Series(tokn[T <= tc - 3600]).groupby(W[T <= tc - 3600]).sum() if (T <= tc - 3600).any() else pd.Series(dtype=float) bmask = S == 1 bw = pd.Series(E[bmask]).groupby(W[bmask]).sum() firstbuy = pd.Series(T[bmask]).groupby(W[bmask]).min() fresh_eth = 0.0; farm_n = 0; known = 0 for wa, e_ in bw.items(): fb, via = fund.get(wa, (np.nan, "")) if pd.notna(fb): known += 1 if firstbuy[wa] - btt([int(fb)])[0] < 3600: fresh_eth += e_ if via in DISP: farm_n += 1 al = sum(1 for wa in bw.index if alum.get(wa, 9e18) < tc) # impact: virtual ETH reserve from recent swaps Pi = P[-31:]; Ei = E[-31:] dl = np.abs(np.diff(np.log(Pi))); ev = 2 * Ei[1:] / np.where(dl > 1e-9, dl, np.nan) Ev = np.nanmedian(ev) if np.isfinite(ev).any() else np.nan # outcomes from entry (tc + DELAY) on te = tc + DELAY if pd.notna(pull_t) and pull_t <= te: rows.append(dict(tok=tok, cp=cpn, alive=0, day=day)); continue ie = np.searchsorted(t, te, side="left") if ie >= len(t): continue pxe = px[ie]; te = t[ie] fut = (t > te) & (t <= te + 7 * 86400) sp = px[fut & sellable] hit = np.nonzero(fut & sellable & (px >= 5 * pxe))[0] t5 = t[hit[0]] if len(hit) else np.nan mx = sp.max() / pxe if len(sp) else 0.0 # value at +24h t24 = tc + 86400 if pd.notna(pull_t) and pull_t <= t24: cap24 = 0.0 elif t24 > DATA_END: cap24 = np.nan else: cap24 = K * px[np.searchsorted(t, t24, side="right") - 1] # value at end of 7d hold (pull = 0) t7 = tc + 7 * 86400 if pd.notna(pull_t) and pull_t <= min(t7, DATA_END): end_x = 0.0 else: end_x = px[np.searchsorted(t, min(t7, DATA_END), side="right") - 1] / pxe cpast = cl_by_cr.get(cr) rows.append(dict( tok=tok, cp=cpn, alive=1, day=day, t0=lr.t0, tc=tc, age_min=(tc - lr.t0) / 60, cap=capc, ethusd=eu, Ev=Ev, hfam=lr.hfam, start=lr.start, crew=crew, cr=cr, close_high=pxc / P.max(), churn=E[last1h].sum() / cap_eth, bs_ratio=(S[last1h] == 1).sum() / max(1, (S[last1h] == -1).sum()), netflow15=(E[last15] * S[last15]).sum() / cap_eth, top10=pos.nlargest(10).sum() / pos.sum() if len(pos) else np.nan, holders=len(pos), hold_growth=(len(pos) + 1) / ((hold1 > 0).sum() + 1), fresh_share=fresh_eth / bw.sum() if bw.sum() > 0 else np.nan, farm_share=farm_n / max(1, len(bw)), fund_known=known / max(1, len(bw)), alumni=al, alumni_share=al / max(1, len(bw)), buyers=len(bw), cr_past_n=nb(cpast, lr.t0 - 1) if cpast is not None else 0, cr_past500=int(nb(past500_cr.get(cr), tc) > 0), root_past500=int(nb(past500_root.get(root), tc) > 0) if isinstance(root, str) else 0, cr_age_h=(lr.t0 - btt([int(cr_fund[0])])[0]) / 3600 if pd.notna(cr_fund[0]) else np.nan, cr_disp=int(cr_fund[1] in DISP), cr_sold=(E[(W == cr) & (S == -1)]).sum() / cap_eth, crew_past1m=int(nb(crew1m.get(crew), tc) > 0) if crew is not None else 0, Y5=int(len(hit) > 0), t5=t5, maxx=mx, cap24=cap24, Y1M=(np.nan if pd.isna(cap24) else int(cap24 >= 1e6)), end_x=end_x, pk7=X7.get(tok, np.nan), pulled=int(pd.notna(pull_t)), pull_min=(pull_t - tc) / 60 if pd.notna(pull_t) else np.nan, )) F = pd.DataFrame(rows) F.to_csv(R + OUTF, index=False) print(F.groupby("cp").agg(n=("tok", "size"), alive=("alive", "sum"), y5=("Y5", "sum"), y1m=("Y1M", "sum")))