Question

我正在尝试使用PyMC3将高斯过程回归函数拟合到一些基本的金融时间序列数据，以便预测接下来几天给定过去价格的“价格”。然而，当我试图从合适的GP中形成预测时，我遇到了问题。

请原谅我，如果其他地方已经涉及过;我已经做了很多挖掘和搜索，但是尽管使用PyMC3进行了GP回归的各种相关教程，但我无法提出解决方案。

通过获取每日价格的自然对数，然后从detrend应用statsmodels.tsa.tsatools，时间序列已经转换为静止（在某种意义上），然后应用In: X[0] Out: array([ 0.02824104, 0.02591076, 0.00990705, 0.00150235, 0.01375847, 0.02963981])。

输入X变量是6维（过去6个价格），这是一个数据点的示例：

In: y[0]

Out: array([ 0.01260668])

关联的y变量是下一天的价格：

import pymc3 as pm

with pm.Model() as gp_fit:

    # Prior beliefs in hyperparameter values (they're Gamma distributed as specified) for Matern 3/2 kernel 
    ρ = pm.Gamma('ρ', 1, 1)
    η = pm.Gamma('η', 1, 1)

    # The input is 6 dimensional hence Matern32(6, ρ)
    K = η * pm.gp.cov.Matern32(6, ρ)

    # Prior belief in mean of the Gaussian Process (it's zero)
    M = pm.gp.mean.Zero()

    # Prior noise belief (Standard Normal, 0 mean, unit variance)
    σ = pm.Normal('σ', 0, 1)

    # Fit to first 20 data points
    y_obs = pm.gp.GP('y_obs', mean_func=M, cov_func=K, sigma=σ, observed={'X':X[:20], 'Y':y[:20]})

    trace = pm.sample(2000, n_init=40000)

我适合GP的代码是：

Auto-assigning NUTS sampler...
Initializing NUTS using ADVI...
Average Loss = -2,456.9:  83%|████████▎ | 33167/40000 [00:38<00:06, 1133.80it/s]
Convergence archived at 33200
Interrupted at 33,200 [83%]: Average Loss = -1,467.9
100%|█████████▉| 2499/2500 [17:29<00:00, 15.29it/s] /home/bango/miniconda3/lib/python3.6/site-packages/pymc3/step_methods/hmc/nuts.py:448: UserWarning:     Chain 0 reached the maximum tree depth. Increase max_treedepth, increase target_accept or reparameterize.
  'reparameterize.' % self._chain_id)
/home/bango/miniconda3/lib/python3.6/site-packages/pymc3/step_methods/hmc/nuts.py:440: UserWarning: The acceptance probability in chain 0 does not match the target. It is 0.18543857929, but should be close to 0.8. Try to increase the number of tuning steps.
  % (self._chain_id, mean_accept, target_accept))
100%|██████████| 2500/2500 [17:29<00:00,  2.38it/s]

它给出了输出（带有一些警告）：

# Get the features of a single data point and rehape to (1,6)
X_test = X[20][None,:]

with gp_fit:
    gp_samples = pm.gp.sample_gp(trace, y_obs, X_test, samples=50, random_seed=666, progressbar=False)

随后，我尝试通过以下方式从拟合的GP中获得预测的y值（给出一些新的观察结果）：

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-29-9eaebf0e139a> in <module>()
      1 with gp_fit:
----> 2     gp_samples = pm.gp.sample_gp(trace, y_obs, Z, samples=50, random_seed=666, progressbar=False)

~/miniconda3/lib/python3.6/site-packages/pymc3/gp/gp.py in sample_gp(trace, gp, X_values, samples, obs_noise, model, random_seed, progressbar)
    139     gp_post = MvNormal.dist(m_post, S_post, shape=Z.shape[0])
    140 
--> 141     samples = [gp_post.random(point=trace[idx]) for idx in indices]
    142 
    143     return np.array(samples)

~/miniconda3/lib/python3.6/site-packages/pymc3/gp/gp.py in <listcomp>(.0)
    139     gp_post = MvNormal.dist(m_post, S_post, shape=Z.shape[0])
    140 
--> 141     samples = [gp_post.random(point=trace[idx]) for idx in indices]
    142 
    143     return np.array(samples)

~/miniconda3/lib/python3.6/site-packages/pymc3/distributions/multivariate.py in random(self, point, size)
    156             mu, cov = draw_values([self.mu, self.cov], point=point)
    157             if mu.shape != cov[0].shape:
--> 158                 raise ValueError("Shapes for mu an cov don't match")
    159 
    160             try:

ValueError: Shapes for mu an cov don't match

但碰到：

$(document).ready(function() {

  //location API

  $("#myweather").load("window",function() {

    $("#location").fadeOut(function() {
      $.ajax({
        async: false,
        url: "https://geoip-db.com/jsonp",
        jsonpCallback: "callback",
        dataType: "jsonp",
        success: function(loc) {
          $("#location").html(loc.city)
          var long = loc.longitude
          var lat = loc.latitude

          $("#result").load("location", function() {
            $("#result").fadeIn(function() {
              $.getJSON("https://api.darksky.net/forecast/15f8bf5641489ec32f66662221933c14/" + lat + long, function(forecast) {
                consolelog(forecast);
              });
            });
          });
        }
      });
    }).fadeIn();
  });
});

我已经挖了一段时间，但目前无法分开错误并纠正它。我不完全确定是什么mu和cov矩阵被引用。

非常感谢您对此的帮助，谢谢！

PyMC3：从高斯过程回归中检索“预测”

0 个答案: