SPB Git

Repository Public

file_code.json 1.4 KB 1 lines json

1{"path":"code/fc_ucsv.py","bytes":1323,"totalLines":29,"from":1,"to":29,"text":"\nimport sys; sys.path.insert(0,\"code\")\nfrom fcsuite import *\ngamma = float(sys.argv[1]); tag=f\"ucsv_g{sys.argv[1]}\"\nN = 40000; rng = np.random.default_rng(12345)\nd = load(); n=len(d); y = d[\"infl\"].values\ntau = rng.normal(y[:4].mean(), 3.0, N)\nh = rng.normal(np.log(np.var(y[:20])*0.8), 1.0, N)\ng = rng.normal(np.log(0.3), 1.0, N)\ntaus=np.zeros(n); ess=np.zeros(n)\nfor t in range(n):\n    h = h + gamma*rng.standard_normal(N); g = g + gamma*rng.standard_normal(N)\n    tau = tau + np.exp(g/2)*rng.standard_normal(N)\n    lw = -0.5*h - 0.5*(y[t]-tau)**2*np.exp(-h)\n    w = np.exp(lw-lw.max()); w/=w.sum()\n    ess[t]=1/np.sum(w**2)\n    taus[t]=np.sum(w*tau)\n    idx = rng.choice(N,N,p=w)             # multinomial resampling every period\n    tau,h,g = tau[idx],h[idx],g[idx]\nrows=[]\nfor hh in H_LIST:\n    yt = target(d,hh).values\n    for T in origins(d):\n        if T+hh>n-1: continue\n        rows.append((str(d.index[T]),hh,\"UCSV\",taus[T],yt[T]))\nout=pd.DataFrame(rows,columns=[\"origin\",\"h\",\"model\",\"fc\",\"actual\"]); out.to_csv(f\"artifacts/forecasts/{tag}.csv\",index=False)\npd.DataFrame({\"period\":d.index.astype(str),\"tau\":taus,\"ess\":ess}).to_csv(f\"artifacts/forecasts/{tag}_tau.csv\",index=False)\nprint(tag,\"min ESS\",ess.min().round(0)); print(out.groupby(\"h\").apply(lambda g_: np.sqrt(np.mean((g_.fc-g_.actual)**2))).round(3))\n","truncated":false}
Files main
Menu

Theme