jackandCursor
3b14bba247
三币过完三滤网只有 1.72 笔/天，攒 30 笔要 17 天。但滑点本来每根都在记，信号 时刻的测量只多回答一个问题：信号那一刻的流动性是否比普通根差。这个问题可以 用 210 天历史离线回答。 关键是必须做匹配对照。信号按 ATR ≥ 8bp 门控，信号根天然比平均根波动大，直接 和全体根比一定会「发现」一个我们自己施加的差异。所以对照组按「同时段 × 同 ATR 十分位」抽取，并排除距信号 48 根内的根（持仓期不独立）。自检 ATR 比值 0.976~1.004，匹配成立。 结果三币一致，方向与担心的相反：信号根成交额是对照的 2.1~2.6 倍、Amihud 非 流动性只有 0.47~0.60 倍、Roll 有效价差三个币都不显著。信号跟在突破后面， 突破自带成交量。所以全体根测出的冲击与价差偏保守而非偏乐观。 唯一真实差异是根内波幅宽 25~28%，但那是波动而非流动性，对应漂移而非冲击， 且可直接当缩放系数：1 秒延迟点漂移上调后仍只占预算 1.4%/3.0%/4.5%。 判读按流动性与波动分两组。初版把 range_bp 当成流动性红旗，会得出「信号时刻 更差、必须等样本」的相反结论——它是波动度量，且 ATR 已匹配。 检验用置换而非 t：成交额跨几个数量级，右尾太重。scipy 未安装，也不宜在采集 运行中动环境，所以用 numpy 自己实现。 Co-authored-by: Cursor <cursoragent@cursor.com>
14 lines
1.2 KiB
CSV
14 lines
1.2 KiB
CSV
sym,proxy,sig_med,ctrl_med,ratio,p,n_sig,n_ctrl
|
|
BTC,vol_usd,5550103.60788,2624836.584895,2.114456816023851,9.999000099990002e-05,91,1820
|
|
BTC,range_bp,14.211131842272723,11.379813132903628,1.248801863115187,0.0007999200079992001,91,1820
|
|
BTC,amihud,0.9320529751601416,1.995826823303005,0.4670009262715666,0.0007999200079992001,91,1820
|
|
BTC,roll_bp,4.779754159181145,5.745083186251997,0.8319730113950504,0.08029197080291971,44,996
|
|
ETH,vol_usd,3126641.7427,1207249.22975,2.589889200714149,9.999000099990002e-05,126,2520
|
|
ETH,range_bp,15.77081223459847,12.35743868645172,1.2762201484267979,0.00029997000299970003,126,2520
|
|
ETH,amihud,2.410291522791945,4.420234301516299,0.5452859188855959,0.0018998100189981002,126,2520
|
|
ETH,roll_bp,6.347567691125725,5.949780489978004,1.0668574583243478,0.42275772422757724,56,1362
|
|
SOL,vol_usd,345692.792,162027.9883,2.1335375179746032,9.999000099990002e-05,145,2900
|
|
SOL,range_bp,15.652030645029573,12.452668963002594,1.25692176444523,0.00019998000199980003,145,2900
|
|
SOL,amihud,20.14951190677575,33.887234963568275,0.594604780485579,0.001999800019998,145,2900
|
|
SOL,roll_bp,5.7402773658108215,6.019208571018624,0.9536598205699658,0.5864413558644136,73,1632
|