Excursions.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "# Overview\n",
        "\n",
        "In this notebook we run three variations on our main models.  The main finding is that these alternative approaches do not markedly improve our results or change our basic findings.\n",
        "\n",
        "1. Macro model with classification target (next rate up or not)\n",
        "\n",
        "2. Text analysis with richer corpus\n",
        "\n",
        "3. Text analysis with noise-based text augmentation.\n"
      ],
      "metadata": {
        "id": "d_r3b7gjkXUS"
      }
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "k1jV3IklNUb9"
      },
      "source": [
        "# **Imports**\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "jKJNCSme3J8U"
      },
      "outputs": [],
      "source": [
        "import numpy as np\n",
        "import pandas as pd\n",
        "\n",
        "%matplotlib inline\n",
        "import matplotlib.pyplot as plt\n",
        "import os\n",
        "import json\n",
        "\n",
        "import tensorflow as tf\n",
        "from tensorflow.keras.preprocessing.text import Tokenizer\n",
        "from tensorflow.keras.preprocessing.sequence import pad_sequences\n",
        "from keras import Sequential\n",
        "from keras.layers import Embedding, LSTM, Dense, Dropout\n",
        "from keras.callbacks import EarlyStopping\n",
        "\n",
        "from sklearn.preprocessing import minmax_scale\n",
        "from sklearn.preprocessing import MinMaxScaler\n",
        "\n",
        "pd.options.display.max_colwidth = 999\n"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5Nq8viF3C4Ad"
      },
      "source": [
        "# **Read and View Data**\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "AwvJe8gh3Vth"
      },
      "outputs": [],
      "source": [
        "df=pd.read_excel('us_ecb_nonumbs_clean_final8.xlsx')"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "TUBETiSa3eiY",
        "outputId": "908b0e0b-cc9d-450b-a935-d2090722c63d"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "<class 'pandas.core.frame.DataFrame'>\n",
            "RangeIndex: 434 entries, 0 to 433\n",
            "Data columns (total 15 columns):\n",
            " #   Column                  Non-Null Count  Dtype  \n",
            "---  ------                  --------------  -----  \n",
            " 0   bank                    434 non-null    object \n",
            " 1   filename                434 non-null    object \n",
            " 2   date_dt                 434 non-null    object \n",
            " 3   chron_order_in_country  434 non-null    int64  \n",
            " 4   rate_curr               434 non-null    float64\n",
            " 5   rate_next               434 non-null    float64\n",
            " 6   rate_change             434 non-null    int64  \n",
            " 7   urate                   434 non-null    float64\n",
            " 8   inf                     434 non-null    float64\n",
            " 9   eff_rate                434 non-null    float64\n",
            " 10  statement               434 non-null    object \n",
            " 11  text_include            434 non-null    int64  \n",
            " 12  nairu                   169 non-null    float64\n",
            " 13  twice_urate_less_nairu  169 non-null    float64\n",
            " 14  minutes                 162 non-null    object \n",
            "dtypes: float64(7), int64(3), object(5)\n",
            "memory usage: 51.0+ KB\n"
          ]
        }
      ],
      "source": [
        "df.info()"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "D9SuoyxbqF7G"
      },
      "outputs": [],
      "source": [
        "#Preprocessing\n",
        "def make_files(data,shuffle):\n",
        "  if shuffle==1:\n",
        "    data=data.sample(frac=1, random_state=42).reset_index()\n",
        "  else:\n",
        "    data=data\n",
        "\n",
        "  cut1=int(.6*data.shape[0])\n",
        "  cut2=int(.8*data.shape[0])\n",
        "\n",
        "  data_train=data.iloc[0:cut1,:].reset_index()\n",
        "  data_val=data.iloc[cut1:cut2,:].reset_index()\n",
        "  data_test=data.iloc[cut2:data.shape[0],:].reset_index()\n",
        "\n",
        "  return data_train, data_val, data_test"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "0tOXmLVXh5DY",
        "outputId": "07b91e23-c2b6-4f01-9152-6b5d55733806"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(101, 17) (34, 17) (34, 17)\n"
          ]
        }
      ],
      "source": [
        "#Files for USFED\n",
        "df_train_us, df_val_us, df_test_us = make_files(df[df.bank=='usfed'], shuffle=1)\n",
        "print(df_train_us.shape, df_val_us.shape, df_test_us.shape)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "VX5JFfsHEVq6"
      },
      "outputs": [],
      "source": [
        "def make_matrixes(df,augment,scaled=0):\n",
        "\n",
        "  if scaled==1 and augment !=\"augment\":\n",
        "    xcols=['urate_std','inf_std','eff_rate_std']\n",
        "    y='rate_next'\n",
        "    #y='rate_curr'\n",
        "    X_matrix=df[xcols].values #x_col_list is a subset of columns in a pandas df\n",
        "    y_col=df[y].values\n",
        "\n",
        "  elif scaled==0:\n",
        "    xcols=['urate','inf','eff_rate']\n",
        "    y='rate_next'\n",
        "    #y='rate_curr'\n",
        "    y_col=df[y].values\n",
        "\n",
        "    #Note: augmentation with noise is only used for training data.\n",
        "    if augment=='augment':\n",
        "      X_matrix = df[xcols].values + tf.random.normal((3,), 0, 0.005, seed=42)\n",
        "\n",
        "    else :\n",
        "      X_matrix = df[xcols].values\n",
        "\n",
        "  print(X_matrix.shape, y_col.shape) #check\n",
        "\n",
        "  return X_matrix, y_col"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "M3_F5XL8laSy",
        "outputId": "40385455-c3b9-4ecf-e4e8-990f124087eb"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(101, 3) (101,)\n",
            "(34, 3) (34,)\n",
            "(34, 3) (34,)\n"
          ]
        }
      ],
      "source": [
        "# Retrieve matrixes for US\n",
        "# Note augmentation can only be used for training data, not validation or test\n",
        "X_train_us, y_train_us = make_matrixes(df_train_us,\"augment\",scaled=0)\n",
        "\n",
        "X_val_us, y_val_us = make_matrixes(df_val_us,\"no aug\",scaled=0)\n",
        "\n",
        "X_test_us, y_test_us = make_matrixes(df_test_us,\"no aug\",scaled=0)"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "X_train_us[0:5]"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "1yYjLB8grxew",
        "outputId": "c6f78077-0d91-41d0-a7cd-44d9a7b6d36d"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<tf.Tensor: shape=(5, 3), dtype=float32, numpy=\n",
              "array([[3.8985963 , 2.0840611 , 2.3966184 ],\n",
              "       [4.698596  , 2.1425312 , 4.466618  ],\n",
              "       [4.7985964 , 1.5805312 , 0.29661837],\n",
              "       [4.8985963 , 2.2796612 , 4.0866184 ],\n",
              "       [3.6985962 , 1.7576312 , 2.3966184 ]], dtype=float32)>"
            ]
          },
          "metadata": {},
          "execution_count": 26
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# **1.  Macro model with classification (up or not)**\n",
        "\n",
        "Test precision was as high as 75%.  However, the test performance was significantly better than the validation performance which suggests the test result is not reliable."
      ],
      "metadata": {
        "id": "qcxh3-aaoXRA"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "def rate_change_var(data):\n",
        "  var=np.where(data['rate_change'] == -1,\n",
        "                               0, data['rate_change'])\n",
        "  return var"
      ],
      "metadata": {
        "id": "8ECep3XjpUbU"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "clf_test_us = np.array(rate_change_var(df_test_us))\n",
        "clf_train_us = np.array(rate_change_var(df_train_us))\n",
        "clf_val_us=np.array(rate_change_var(df_val_us))"
      ],
      "metadata": {
        "id": "pJsAlw_RpUei"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "print(clf_test_us.shape, clf_train_us.shape, clf_val_us.shape)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "P5UgIjaTrERz",
        "outputId": "7bcb5962-bf8f-4ca3-df74-7556136d151a"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(34,) (101,) (34,)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "model_mac_precision = tf.keras.Sequential([\n",
        "    tf.keras.layers.InputLayer(input_shape=(3,)),\n",
        "    tf.keras.layers.Dense(480, activation='relu'),\n",
        "    tf.keras.layers.Dense(464, activation='relu'),\n",
        "    tf.keras.layers.Dense(1, activation='sigmoid')])"
      ],
      "metadata": {
        "id": "nSQUGSyJnZcw"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "tf.keras.backend.clear_session()\n",
        "np.random.seed(42)\n",
        "tf.random.set_seed(42)\n",
        "model_mac_precision.compile(loss='binary_crossentropy',\n",
        "              optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),\n",
        "              metrics=[tf.keras.metrics.Precision()])"
      ],
      "metadata": {
        "id": "W-pMslo8nZgA"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "num_epochs = 2500\n",
        "early_stop_clf = EarlyStopping(monitor='val_precision',\n",
        "                           mode='max', patience=300, verbose=False)\n",
        "history = model_mac_precision.fit(X_train_us, clf_train_us,\n",
        "                    epochs=num_epochs,\n",
        "                    callbacks=[early_stop_clf],\n",
        "                    verbose=False,\n",
        "                    validation_data=(X_val_us, clf_val_us))"
      ],
      "metadata": {
        "id": "7psX4DLwnZlG"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "precision = history.history['precision']\n",
        "val_precision = history.history['val_precision']\n",
        "epochs=range(len(precision))\n",
        "plt.plot(epochs, precision, 'r')\n",
        "plt.plot(epochs, val_precision, 'b')\n",
        "plt.title('Training and validation Precision')\n",
        "plt.ylim((0,1.5))\n",
        "plt.figure()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 317
        },
        "id": "ku_Annnsnjfe",
        "outputId": "4a7fcf6a-bd6b-4d53-adea-b8bd163ac2bb"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 0 Axes>"
            ]
          },
          "metadata": {},
          "execution_count": 37
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 0 Axes>"
            ]
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "model_mac_precision.evaluate(X_test_us, clf_test_us)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "VlJNmmhbnji6",
        "outputId": "d2c4b234-f66e-47c1-9fde-778cbcf4a9c9"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "2/2 [==============================] - 0s 8ms/step - loss: 0.2499 - precision: 0.7500\n"
          ]
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[0.24986295402050018, 0.75]"
            ]
          },
          "metadata": {},
          "execution_count": 38
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "KtWZGAEhurwC"
      },
      "source": [
        "# **2.  Richer Text**\n",
        "\n",
        "Our main dataset consists of short and concise statements which indicate the final considerations for the policy choice.  Most researchers, however, have used the much more extensive meeting minutes.  These may give a better window into the uncertainty, optimism and pessimism of the monetary policy committee.  \n",
        "On the other hand, these long statements may difficult for an LSTM model, since LSTM models do not use transformers.\n",
        "\n",
        "To determine the overall impact, we ran the richer text data on our simple LSTM and compared it to the original results.  "
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#Statistics\n",
        "approx_length = df_train_us['minutes'].str.split(\" \").str.len()\n",
        "approx_length.describe()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "SWg8jETmuN86",
        "outputId": "fb4df90d-e2ee-480b-8e8c-20ac4b27bf82"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "count      98.000000\n",
              "mean     2550.163265\n",
              "std       765.612519\n",
              "min        89.000000\n",
              "25%      2147.750000\n",
              "50%      2992.500000\n",
              "75%      3033.750000\n",
              "max      3131.000000\n",
              "Name: minutes, dtype: float64"
            ]
          },
          "metadata": {},
          "execution_count": 39
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "_v6l6_9wySCE"
      },
      "source": [
        "## Prepare data for embeddings\n",
        "We must drop the na values for minutes and create the padded sequences.\n"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "df_train_us=df_train_us.dropna()\n",
        "df_val_us=df_val_us.dropna()\n",
        "df_test_us=df_test_us.dropna()\n",
        "\n",
        "#check\n",
        "df_train_us.info()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "EGhVmGxmux-M",
        "outputId": "8ad86ba7-783c-4c51-931d-4e900251d89d"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "<class 'pandas.core.frame.DataFrame'>\n",
            "Int64Index: 98 entries, 0 to 100\n",
            "Data columns (total 17 columns):\n",
            " #   Column                  Non-Null Count  Dtype  \n",
            "---  ------                  --------------  -----  \n",
            " 0   level_0                 98 non-null     int64  \n",
            " 1   index                   98 non-null     int64  \n",
            " 2   bank                    98 non-null     object \n",
            " 3   filename                98 non-null     object \n",
            " 4   date_dt                 98 non-null     object \n",
            " 5   chron_order_in_country  98 non-null     int64  \n",
            " 6   rate_curr               98 non-null     float64\n",
            " 7   rate_next               98 non-null     float64\n",
            " 8   rate_change             98 non-null     int64  \n",
            " 9   urate                   98 non-null     float64\n",
            " 10  inf                     98 non-null     float64\n",
            " 11  eff_rate                98 non-null     float64\n",
            " 12  statement               98 non-null     object \n",
            " 13  text_include            98 non-null     int64  \n",
            " 14  nairu                   98 non-null     float64\n",
            " 15  twice_urate_less_nairu  98 non-null     float64\n",
            " 16  minutes                 98 non-null     object \n",
            "dtypes: float64(7), int64(5), object(5)\n",
            "memory usage: 13.8+ KB\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#We will also need to reset the y values and labels\n",
        "y_train_us=np.array(df_train_us.rate_next)\n",
        "y_test_us=np.array(df_test_us.rate_next)\n",
        "y_val_us=np.array(df_val_us.rate_next)"
      ],
      "metadata": {
        "id": "0RxM8J1NvSHU"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def prep_text(texttrain, textval, texttest):\n",
        "  tr_sentences=np.array(texttrain)\n",
        "  val_sentences=np.array(textval)\n",
        "  test_sentences=np.array(texttest)\n",
        "  return tr_sentences, val_sentences, test_sentences"
      ],
      "metadata": {
        "id": "cQpP5mNJPh2H"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "D7Cey_xKL5ld"
      },
      "outputs": [],
      "source": [
        "#Convert minutes to numpy arrays - US\n",
        "training_sentences_us, val_sentences_us, testing_sentences_us = prep_text(df_train_us.minutes,\\\n",
        "                                                                          df_val_us.minutes,\\\n",
        "                                                                          df_test_us.minutes)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "p3sV3KW7_MRj",
        "outputId": "85c3c3a5-276c-43c2-c81b-5908f0b73830"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(98,) (32,) (32,)\n"
          ]
        }
      ],
      "source": [
        "#Convert rename\n",
        "# Continuous target variable\n",
        "training_labels_final_us, val_labels_final_us, test_labels_final_us = prep_text(y_train_us, y_val_us, y_test_us)\n",
        "\n",
        "print(training_labels_final_us.shape, val_labels_final_us.shape, test_labels_final_us.shape)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "1dJsVBZp5wgc"
      },
      "outputs": [],
      "source": [
        "#Create embeddings based on the corpus\n",
        "\n",
        "max_length = 3000 #Number of tokens per statement\n",
        "trunc_type = 'post'\n",
        "padding_type = 'post'\n",
        "\n",
        "vocab_size = 6000\n",
        "oov_tok = '<OOV>'\n",
        "embedding_dim = 16"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "wHFQvZHq4-J0"
      },
      "outputs": [],
      "source": [
        "#Apply tokenizer which translates words into numbers\n",
        "tokenizer = Tokenizer(num_words=vocab_size, oov_token=oov_tok)\n",
        "tokenizer.fit_on_texts(training_sentences_us)\n",
        "word_index = tokenizer.word_index"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "lahsZxxJ6jVt",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "6323b529-c722-4e2a-c69f-e9cbf43b0868"
      },
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[57, 104, 1388, 4127]"
            ]
          },
          "metadata": {},
          "execution_count": 96
        }
      ],
      "source": [
        "#Find the indexes for words we could not find before.\n",
        "check_list=['staff','housing', 'macroeconomic', 'war']\n",
        "\n",
        "[word_index[i] for i in check_list]"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "len(word_index) #there are a lot more words - almost 5 times more"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "pDXpqd4kYQz2",
        "outputId": "9dc4d5b7-b78a-4dce-8832-4badaa6a4a99"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "5940"
            ]
          },
          "metadata": {},
          "execution_count": 85
        }
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "y9PFBhRj6nbQ"
      },
      "outputs": [],
      "source": [
        "#pad sequences\n",
        "sequences_us = tokenizer.texts_to_sequences(training_sentences_us)\n",
        "padded_us = pad_sequences(sequences_us, maxlen=max_length, truncating=trunc_type)\n",
        "\n",
        "val_sequences_us = tokenizer.texts_to_sequences(val_sentences_us)\n",
        "val_padded_us = pad_sequences(val_sequences_us, maxlen=max_length)\n",
        "\n",
        "test_sequences_us = tokenizer.texts_to_sequences(testing_sentences_us)\n",
        "test_padded_us = pad_sequences(test_sequences_us, maxlen=max_length)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "XyORPwqpvBpD"
      },
      "source": [
        "Let's take a look at these sequences arrays."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "EK-QzS4buLo8",
        "outputId": "cc7cb02a-6a31-4e48-d689-196d1ccedf62"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(98, 3000) (32, 3000) (32, 3000)\n"
          ]
        }
      ],
      "source": [
        "print(padded_us.shape, val_padded_us.shape, test_padded_us.shape)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "9Y9Hqj3EWqsW",
        "outputId": "ccb0788f-2443-41a5-becf-322fd0666475"
      },
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "array([[ 135,    5,  155, ...,  648,  269,   59],\n",
              "       [   0,    0,    0, ...,  567,  629,  577],\n",
              "       [ 135,    5,  155, ...,   52,  187,   26],\n",
              "       ...,\n",
              "       [   0,    0,    0, ...,    4,  636, 5907],\n",
              "       [ 135,    5,  155, ...,  774,  212,  140],\n",
              "       [   0,   13,  180, ...,   21,   67, 5940]], dtype=int32)"
            ]
          },
          "metadata": {},
          "execution_count": 99
        }
      ],
      "source": [
        "padded_us"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "### A simple LSTM.  \n",
        "\n",
        "Previously, the test rmse was .52.  With the new data, it was about .8 - much worse.  We also tried doubling the size of the embedding dimensions to accomodate a richer dataset, which did not help. Overall, this model trained very well but was over-fit to the data and did poorly in both sample and validation tests.  It is likey that a different tool was needed for the job of analyzing these long texts."
      ],
      "metadata": {
        "id": "NdJ8x2uKFJGi"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "#Now run the model!\n",
        "model_lstm = tf.keras.Sequential([\n",
        "    tf.keras.layers.Embedding(vocab_size, embedding_dim, input_length=max_length),\n",
        "    tf.keras.layers.LSTM(embedding_dim),\n",
        "    tf.keras.layers.Dense(480,activation='relu'),\n",
        "    tf.keras.layers.Dense(464,activation='relu'),\n",
        "    tf.keras.layers.Dense(1)\n",
        "])"
      ],
      "metadata": {
        "id": "Dm3mJ3iK093K"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "tf.keras.backend.clear_session()\n",
        "np.random.seed(42)\n",
        "tf.random.set_seed(42)\n",
        "model_lstm.compile(loss='mse',\n",
        "              optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),\n",
        "              metrics=[tf.keras.metrics.RootMeanSquaredError()])"
      ],
      "metadata": {
        "id": "khJCAdsYFRji"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "num_epochs = 2500\n",
        "early_stop = EarlyStopping(monitor='val_root_mean_squared_error',\n",
        "                           mode='min', patience=300, verbose=False)\n",
        "history = model_lstm.fit(padded_us, y_train_us,\n",
        "                    epochs=num_epochs, verbose=False, callbacks=[early_stop],\n",
        "                    validation_data=(val_padded_us, y_val_us))"
      ],
      "metadata": {
        "id": "-ihWrv7uFXDy"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#@title\n",
        "rmse = history.history['root_mean_squared_error']\n",
        "val_rmse = history.history['val_root_mean_squared_error']\n",
        "epochs=range(len(rmse))\n",
        "plt.plot(epochs, rmse, 'r')\n",
        "plt.plot(epochs, val_rmse, 'b')\n",
        "plt.title('Training and validation RMSE - Simple LSTM, US')\n",
        "plt.ylim((0,1.5))\n",
        "plt.figure()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 317
        },
        "id": "9biNc1BiFZ5P",
        "outputId": "174d25cc-4525-4fde-eb84-36167970aace"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 0 Axes>"
            ]
          },
          "metadata": {},
          "execution_count": 103
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 0 Axes>"
            ]
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "model_lstm_eval=model_lstm.evaluate(test_padded_us, y_test_us)\n",
        "model_lstm_eval"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "x-rHEQ-cFdVL",
        "outputId": "502f3663-6230-46af-fd01-b5102997e4b9"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "1/1 [==============================] - 0s 77ms/step - loss: 0.6896 - root_mean_squared_error: 0.8304\n"
          ]
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[0.6896129250526428, 0.8304293155670166]"
            ]
          },
          "metadata": {},
          "execution_count": 104
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# **3.  Text augmentation with noise**\n",
        "\n",
        "Result: Test rmse ~.8, higher than original data.\n",
        "\n",
        "We used the Text Attack package to make small perturbations to the post-meeting (shorter) statements.  \n",
        "\n",
        "According to research, noise can force an nlp model to train over a wider range of inputs, increasing its robustness and test performance.  Furthermore the perturbed texts can be appended to the training dataset, increasing its size.  \n",
        "\n",
        "The package offers several augmentation tools and we chose the WordNet Augmenter.  This tool replaces random words in each sample with related in WordNet.  Sometimes it gets teh sense of the word right, other times it substitutes a word that is synonymous only in a sense not used in the text (e.g., \"energy prices\" becomes \"vigor prices\").\n",
        "\n",
        "**The textattack package takes a long time to run even on a small number of samples and GPU.**\n",
        "\n",
        "For this reason we are going to show the code we used to generate the text and read in the data, rather than running the code from this notebook."
      ],
      "metadata": {
        "id": "6PbOd23o3vj5"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "## Set up\n",
        "#! pip install textattack\n",
        "#import textattack\n",
        "#from textattack.augmentation import WordNetAugmenter\n",
        "\n",
        "## Indicate the original datas is not augmented.\n",
        "#df_train_us['text_aug']=\"no\"\n",
        "\n",
        "##Given the time reqiured to run even one sample,choose 20 relatively short ones.\n",
        "#include=[0,1,3,7,10,14,16,20,21,22,23,26,29,30,35,36,38,39,43,65]\n",
        "\n",
        "## Create the dataset that will be augemented\n",
        "#train_aug = df_train_us.iloc[include, :]\n",
        "\n",
        "## Specify parameters\n",
        "#aug = WordNetAugmenter(pct_words_to_swap = 0.1, transformations_per_example = 1,\n",
        "                      #fast_augment=True)\n",
        "\n",
        "## Augment\n",
        "#train_aug['statement'] = train_aug['statement'].apply(lambda x: str(aug.augment(x)))\n",
        "\n",
        "##Indicate new data is augmented.\n",
        "#train_aug['text_aug']=train_aug.text_aug.str.replace('no','yes')\n",
        "\n",
        "##Add new samples to training data\n",
        "#df_train_us=df_train_us.append(train_aug)"
      ],
      "metadata": {
        "id": "1GA03ZRt3uL7"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "Now that we have some augmented training data we can read it in.  The file includes the 20 augmented samples and their original counterparts, along with samples that were not perturbed, bringing up the number of observations from 101 to 121.\n",
        "\n",
        "The test and validation sets are unchanged."
      ],
      "metadata": {
        "id": "bXDqGrd1-e3b"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# Now that we have created some augmented text let's read it in.\n",
        "#\n",
        "df_train_us=pd.read_excel('/content/df_train_us_with_aug2.xlsx')"
      ],
      "metadata": {
        "id": "a2cCVK503ueU"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "df_train_us.shape"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "3iSLQIJJN64m",
        "outputId": "3541a0a7-9b49-4a41-fede-55fab2c95cab"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(121, 13)"
            ]
          },
          "metadata": {},
          "execution_count": 132
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#Examples of perturbed text\n",
        "df_train_us.loc[111,'statement']"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 196
        },
        "id": "u7-3yX7n3uhe",
        "outputId": "60445d20-0abb-47ab-dbb0-614b362945bb"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "\"indicates proletariat mart remains warm economic activity uprise moderate rate. job gains solid average recent months unemployment rate remained low. although growth household spending piece earlier year growth business fixed investment soft. month basis overall inflation inflation items food energy range. market based measures inflation compensation remain grim; survey based measures longer term inflation expectations little changed. consistent statutory mandate committee seeks foster maximum employment price stability. light implications global developments economic outlook well muted inflation pressures committee decided lower target range federal funds rate. action supports committee's view sustained expansion economic activity strong labor market term inflation near committee's symmetric   objective likely outcomes uncertainties outlook remain. committee contemplates future path target tramp federal funds rate preserve monitor implications incoming information economic outlook act appropriate sustain expansion strong labor market inflation near symmetric   objective. determining timing size future adjustments target range federal funds rate committee appraise realized expected economic conditions comparative maximum employment objective symmetric   inflation objective. assessment take account wide array information including measures labor market conditions indicators inflation pressures inflation expectations translate financial international developments. committee conclude reduction aggregate securities arrest organization open market account two months early previously indicated. rosengren pet contact maintain target range federal funds rate.\""
            ],
            "application/vnd.google.colaboratory.intrinsic+json": {
              "type": "string"
            }
          },
          "metadata": {},
          "execution_count": 133
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#Original version\n",
        "df_train_us.loc[112,'statement']"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 196
        },
        "id": "7EPb-fdlN9Bz",
        "outputId": "9ae18f61-aa6a-4928-f681-b1460a5e96fb"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "\"indicates labor market remains strong economic activity rising moderate rate. job gains solid average recent months unemployment rate remained low. although growth household spending picked earlier year growth business fixed investment soft. month basis overall inflation inflation items food energy running. market based measures inflation compensation remain low; survey based measures longer term inflation expectations little changed. consistent statutory mandate committee seeks foster maximum employment price stability. light implications global developments economic outlook well muted inflation pressures committee decided lower target range federal funds rate. action supports committee's view sustained expansion economic activity strong labor market conditions inflation near committee's symmetric   objective likely outcomes uncertainties outlook remain. committee contemplates future path target range federal funds rate continue monitor implications incoming information economic outlook act appropriate sustain expansion strong labor market inflation near symmetric   objective. determining timing size future adjustments target range federal funds rate committee assess realized expected economic conditions relative maximum employment objective symmetric   inflation objective. assessment take account wide range information including measures labor market conditions indicators inflation pressures inflation expectations readings financial international developments. committee conclude reduction aggregate securities holdings system open market account two months earlier previously indicated. rosengren preferred meeting maintain target range federal funds rate.\""
            ],
            "application/vnd.google.colaboratory.intrinsic+json": {
              "type": "string"
            }
          },
          "metadata": {},
          "execution_count": 3
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "Perturbed sentence: indicates proletariat mart remains warm economic activity uprise moderate rate.\n",
        "\n",
        "Original:  indicates labor market remains strong economic activity rising moderate rate."
      ],
      "metadata": {
        "id": "Y55Jo-LuAGfL"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "#Now feed the new data into our functions for formatting.\n",
        "\n",
        "#We will also need to reset the y values and labels since we have more cases.\n",
        "y_train_us=np.array(df_train_us.rate_next)\n",
        "#confirm new shape\n",
        "\n",
        "y_train_us.shape"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "FZYrlqZKPK9u",
        "outputId": "b8f529a3-9404-49ed-e98b-e2aaf6838359"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(121,)"
            ]
          },
          "metadata": {},
          "execution_count": 134
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#In the prior excursion we looked at the (long) minutes;\n",
        "#change text to 'statements'\n",
        "training_sentences_us, val_sentences_us, testing_sentences_us = prep_text(df_train_us.statement,\\\n",
        "                                                                          df_val_us.statement,\\\n",
        "                                                                          df_test_us.statement)"
      ],
      "metadata": {
        "id": "VtlqMMpGPO7J"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#Ensure all the target variables are in the right format and shape\n",
        "training_labels_final_us, val_labels_final_us, test_labels_final_us = prep_text(y_train_us, y_val_us, y_test_us)\n"
      ],
      "metadata": {
        "id": "5eReGpyOPO_J"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "print(training_labels_final_us.shape,val_labels_final_us.shape,test_labels_final_us.shape)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "vIvp2Kt9PPG1",
        "outputId": "7165e73a-bcc9-4d3b-c585-154ed6cb9a0f"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(121,) (32,) (32,)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#Create new embeddings based on the corpus\n",
        "\n",
        "max_length = 1200 #Number of tokens per statement\n",
        "trunc_type = 'post'\n",
        "padding_type = 'post'\n",
        "\n",
        "vocab_size = 1400\n",
        "oov_tok = '<OOV>'\n",
        "embedding_dim = 16"
      ],
      "metadata": {
        "id": "HyatdLCi3uoJ"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#Apply tokenizer which translates words into numbers\n",
        "tokenizer = Tokenizer(num_words=vocab_size, oov_token=oov_tok)\n",
        "tokenizer.fit_on_texts(training_sentences_us)\n",
        "word_index = tokenizer.word_index"
      ],
      "metadata": {
        "id": "gSDEM8sDSe5n"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#Find the indexes for some new words\n",
        "check_list=['mart','drudge', 'mate', 'proletariat','pet']\n",
        "\n",
        "[word_index[i] for i in check_list]"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "wekBbX2eSoYx",
        "outputId": "9b470c7e-7a99-45a9-cb99-a4b0fdc67d73"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[1343, 858, 980, 1342, 1353]"
            ]
          },
          "metadata": {},
          "execution_count": 157
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "len(word_index) #About 200 more words than in the original vocab"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "S0hgxoClSolP",
        "outputId": "b6f22176-19b3-47d8-daf1-3d8cf85ef906"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "1370"
            ]
          },
          "metadata": {},
          "execution_count": 151
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "#pad sequences\n",
        "sequences_us = tokenizer.texts_to_sequences(training_sentences_us)\n",
        "padded_us = pad_sequences(sequences_us, maxlen=max_length, truncating=trunc_type)\n",
        "\n",
        "val_sequences_us = tokenizer.texts_to_sequences(val_sentences_us)\n",
        "val_padded_us = pad_sequences(val_sequences_us, maxlen=max_length)\n",
        "\n",
        "test_sequences_us = tokenizer.texts_to_sequences(testing_sentences_us)\n",
        "test_padded_us = pad_sequences(test_sequences_us, maxlen=max_length)"
      ],
      "metadata": {
        "id": "EE5eY0E3VEyS"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#Now run the model!\n",
        "model_lstm = tf.keras.Sequential([\n",
        "    tf.keras.layers.Embedding(vocab_size, embedding_dim, input_length=max_length),\n",
        "    tf.keras.layers.LSTM(embedding_dim),\n",
        "    tf.keras.layers.Dense(480,activation='relu'),\n",
        "    tf.keras.layers.Dense(464,activation='relu'),\n",
        "    tf.keras.layers.Dense(1)\n",
        "])"
      ],
      "metadata": {
        "id": "FjWGgFtXVvbW"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "tf.keras.backend.clear_session()\n",
        "np.random.seed(42)\n",
        "tf.random.set_seed(42)\n",
        "model_lstm.compile(loss='mse',\n",
        "              optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),\n",
        "              metrics=[tf.keras.metrics.RootMeanSquaredError()])"
      ],
      "metadata": {
        "id": "drbXmaLo3url"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "num_epochs = 2500\n",
        "early_stop = EarlyStopping(monitor='val_root_mean_squared_error',\n",
        "                           mode='min', patience=300, verbose=False)\n",
        "history = model_lstm.fit(padded_us, y_train_us,\n",
        "                    epochs=num_epochs, verbose=False, callbacks=[early_stop],\n",
        "                    validation_data=(val_padded_us, y_val_us))"
      ],
      "metadata": {
        "id": "UtbFe92nVD4s"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#@title\n",
        "rmse = history.history['root_mean_squared_error']\n",
        "val_rmse = history.history['val_root_mean_squared_error']\n",
        "epochs=range(len(rmse))\n",
        "plt.plot(epochs, rmse, 'r')\n",
        "plt.plot(epochs, val_rmse, 'b')\n",
        "plt.title('Training and validation RMSE - Simple LSTM, US, Perturbed')\n",
        "plt.ylim((0,1.5))\n",
        "plt.figure()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 317
        },
        "id": "0bNon96UVMKB",
        "outputId": "d8930cd8-5853-4ef7-c2fc-c0f03a4684af"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 0 Axes>"
            ]
          },
          "metadata": {},
          "execution_count": 162
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 0 Axes>"
            ]
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "model_lstm_eval=model_lstm.evaluate(test_padded_us, y_test_us)\n",
        "model_lstm_eval"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "nlzZFYVHVD8n",
        "outputId": "9e85ef63-40f3-4781-9b24-a3fee1fe9935"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "1/1 [==============================] - 0s 44ms/step - loss: 0.5468 - root_mean_squared_error: 0.7395\n"
          ]
        },
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[0.5468456745147705, 0.7394901514053345]"
            ]
          },
          "metadata": {},
          "execution_count": 163
        }
      ]
    }
  ],
  "metadata": {
    "colab": {
      "provenance": [],
      "collapsed_sections": [
        "KtWZGAEhurwC",
        "_v6l6_9wySCE",
        "NdJ8x2uKFJGi",
        "6PbOd23o3vj5"
      ]
    },
    "gpuClass": "standard",
    "kernelspec": {
      "display_name": "Python 3 (ipykernel)",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.9.7"
    },
    "accelerator": "GPU"
  },
  "nbformat": 4,
  "nbformat_minor": 0
}