test_refined_target_decoding.py

import polars as pl
import numpy as np
import random

# === Load your refined file ===
df = pl.read_csv("datasets/refined/1h/ADAUSDT_1h_refined.csv")

# === Constants ===
VOLUME_K = 0.2
NUM_ROWS = df.height
WINDOW_SIZE = 30

# === Choose a random starting index safely (must be > 0 and leave 10 rows ahead) ===
start_index = random.randint(1, NUM_ROWS - WINDOW_SIZE - 1)
end_index = start_index + WINDOW_SIZE

print(f"\n🔀 Decoding rows {start_index} to {end_index - 1}:")

# === Iterate over the selected range ===
for i in range(start_index, end_index):
    row_i = df[i]
    row_prev = df[i - 1]

    timestamp = row_i["timestamp"].item()

    prev_close = row_prev["original_close"].item()
    prev_vals = {
        "volume": row_prev["original_volume"].item(),
        "quote_asset_volume": row_prev["original_quote_asset_volume"].item(),
        "number_of_trades": row_prev["original_number_of_trades"].item(),
        "taker_buy_base_asset_volume": row_prev["original_taker_buy_base_asset_volume"].item(),
        "taker_buy_quote_asset_volume": row_prev["original_taker_buy_quote_asset_volume"].item(),
    }

    # === Decode target values from row i ===
    decoded = {}

    # OHLC decoding
    for field in ["open", "high", "low", "close"]:
        target_val = row_i[f"target_{field}"].item()
        decoded[field] = prev_close * np.exp(target_val)

    # Volume-related decoding
    for field in ["volume", "quote_asset_volume", "number_of_trades",
                  "taker_buy_base_asset_volume", "taker_buy_quote_asset_volume"]:
        target_val = row_i[f"target_{field}"].item()
        log_ret = np.arctanh(np.clip(target_val, -0.999999, 0.999999)) / VOLUME_K
        decoded[field] = prev_vals[field] * np.exp(log_ret)

    # === Print decoded vs original comparison ===
    print(f"\n🧪 Decoding check for row {i} @ {timestamp}:")
    for field in decoded:
        original_val = row_i[f"original_{field}"].item()
        decoded_val = decoded[field]
        delta = abs(original_val - decoded_val)
        print(f"{field:<35} original: {original_val:>15.8f}   decoded: {decoded_val:>15.8f}   Δ = {delta:.8f}")