benchmark/plot_forecasts.py

#!/usr/bin/env python3
"""Forecast fan-chart gallery, one facet per model.

Rows = assets, cols = models; per timeframe one PNG. Each facet: recent
real series (grey), the actual future (black), the model's median forecast
(dashed) and its quantile bands (10-90% and 30-70% shaded; for a 13-level
model the trained 2.5-97.5% tail band is shaded very lightly underneath).
Quantile indices are taken from each dump's own `levels` array.

Usage: zoo_plot2.py <tf>   (reads data/forecast-dumps/plotdump-<tag>-<tf>.npz)
"""
import sys
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np

TAGS = [("yv02",     "Yumoto v0.2",   "#0072B2"),
        ("yv01",     "Yumoto v0.1",   "#56B4E9"),
        ("toto25b",  "Toto-2.0-2.5B", "#E69F00"),
        ("tirex2",   "TiRex-2",       "#009E73"),
        ("timesfm3", "TimesFM-3.0",   "#CC79A7")]
TF_TITLE = {"5m": "5-minute", "1h": "hourly", "4h": "4-hourly", "1d": "daily"}

tf = sys.argv[1]
dumps = {}
for tag, _, _ in TAGS:
    try:
        dumps[tag] = np.load(f"data/forecast-dumps/plotdump-{tag}-{tf}.npz",
                             allow_pickle=True)
    except FileNotFoundError:
        pass
ref = next(iter(dumps.values()))
assets = [str(x) for x in ref["assets"]]

def q_at(d, name, level):
    lv = np.asarray(d["levels"], float)
    i = int(np.argmin(np.abs(lv - level)))
    assert abs(lv[i] - level) < 1e-6, (level, lv)
    return np.asarray(d[f"q_{name}"])[:, i]

fig, axes = plt.subplots(len(assets), len(TAGS),
                         figsize=(3.6 * len(TAGS), 2.7 * len(assets)),
                         squeeze=False, dpi=150)
for r, name in enumerate(assets):
    for c, (tag, label, color) in enumerate(TAGS):
        ax = axes[r][c]
        if tag not in dumps:
            ax.axis("off"); continue
        d = dumps[tag]
        ctx = np.asarray(d[f"ctx_{name}"], float)
        act = np.asarray(d[f"act_{name}"], float)
        H = act.size
        # show a context tail ~1.5x the horizon so the forecast has room
        keep = min(ctx.size, int(1.5 * H))
        xs_c = np.arange(-keep, 0); xs_f = np.arange(H)
        lv = np.asarray(d["levels"], float)
        if lv.min() < 0.05:  # 13-level model: show the trained tail band
            ax.fill_between(xs_f, q_at(d, name, 0.025), q_at(d, name, 0.975),
                            color=color, alpha=0.10, lw=0)
        ax.fill_between(xs_f, q_at(d, name, 0.1), q_at(d, name, 0.9),
                        color=color, alpha=0.22, lw=0)
        ax.fill_between(xs_f, q_at(d, name, 0.3), q_at(d, name, 0.7),
                        color=color, alpha=0.38, lw=0)
        ax.plot(xs_c, ctx[-keep:], color="#999999", lw=0.9)
        ax.plot(xs_f, act, color="#111111", lw=1.1)
        ax.plot(xs_f, q_at(d, name, 0.5), color=color, lw=1.2, ls="--")
        ax.axvline(0, color="#bbbbbb", lw=0.7)
        lo = min(ctx[-keep:].min(), act.min()); hi = max(ctx[-keep:].max(), act.max())
        pad = 0.35 * (hi - lo + 1e-12)
        ax.set_ylim(lo - pad, hi + pad)
        ax.tick_params(labelsize=6.5)
        for s in ("top", "right"): ax.spines[s].set_visible(False)
        if r == 0: ax.set_title(label, fontsize=10, color=color)
        if c == 0: ax.set_ylabel(name, fontsize=9)
        if r == len(assets) - 1: ax.set_xlabel("steps from forecast origin", fontsize=7.5)
fig.suptitle(f"{TF_TITLE[tf]} candles — held-out forecasts "
             "(grey: context · black: what actually happened · "
             "dashed: median forecast · shading: quantile bands)",
             fontsize=11, y=1.005)
fig.tight_layout()
fig.savefig(f"plots/forecasts/{tf}.png", bbox_inches="tight")
print("wrote", f"plots/forecasts/{tf}.png")