Is there any certified case that timesfm could predict the trend of a stock?

[JackeyLee007]

I tried it with China's A-share 000001 and found that it was not accurate. Is there anything that I used improperly?

The blue-grey line is real data, and the red line is predicted value. I used the value from 20210101 to 20210430 to predict the value of the next 32 days.

import timesfm
import os
import numpy as np
import akshare as ak
import pandas as pd

stHist = ak.stock_zh_a_hist('000001', 'daily', '20210101', '20210430')
incVals=range(1, len(stHist)+1)
stHist['unique_id'] = incVals
stHist = stHist.rename(columns={'日期':'ds','收盘':'value'})
stHist['ds'] = pd.to_datetime(stHist['ds'])
print(stHist)
tfm = timesfm.TimesFm(
    context_len=128,
    horizon_len=32,
    input_patch_len=32,
    output_patch_len=128,
    num_layers=20,
    model_dims=1280,
    backend='cpu',
)
tfm.load_from_checkpoint(repo_id="google/timesfm-1.0-200m")


forecast_df = tfm.forecast_on_df(
        inputs=stHist,
        freq='D',
        value_name='value',
        num_jobs=1)

[Arata-Nakajima]

I did some modifications to predict next day price only. I run the inference every day instead of predicting multiple days ahead. So far it's accuracy is 67% (14/21) for up or down predictions. I find it very satisfactory. At least my investment performance increased a lot. You may need to adjust parameters in the code (shift_d, discard_th) every day for better accuracy.

Following code works in Google Colab

! git clone https://github.com/google-research/timesfm.git
%cd timesfm
!pip install -e .
!pip install utilsforecast
!pip install transformers accelerate bitsandbytes
!huggingface-cli login --token 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
#!pip install --upgrade numpy
!pip install numpy

import datetime
import yfinance as yf
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import timesfm
from timesfm import TimesFm

# 調整が必要なパラメータ(シフト量と抑制閾値はほぼ毎回調整が必要です)
# フラット抑制を高めると精度は上がるが、horizon(予測期間)が短くなる
date_m = 7 # month
date_d = 11 # day
shift_d = 7 # 時間軸のシフト量
discard_th = 0.00 # フラット抑制閾値(この変動率以下の予測値は切り捨て)
codelist = ["2558.T"] # 銘柄コード
ohlc = "Adj Close" # 4値

# 現在値(手動設定, 必要があれば)
date_str1 = "2024-7-10"
date_str2 = "2024-7-11"
prices1 = [24240.0, 24285.0, 24225.0, 25870.0]
prices2 = [24430.0, 24500.0, 24420.0, 26130.0]
bk_str = "gpu" # Backend

ohlc2i = {"Open": 0, "High": 1, "Low": 2, "Adj Close": 3}
cur_price1 = prices1[ohlc2i[ohlc]]
cur_price2 = prices2[ohlc2i[ohlc]]

start = datetime.date(2020, 1, 1)
end_t = datetime.date(2024, date_m, date_d )
#end_t = datetime.date.today()

data_train = yf.download(codelist, start = start, end = end_t)
data_all = yf.download(codelist, start = start, end = end_t)
print( data_train )

data_all = data_all[ohlc].dropna()  #欠損値を除去

if data_all.empty:
    raise ValueError("データが空です。期間を変更して再度試してください。")
data_all.loc[date_str1] = cur_price1
data_all.loc[date_str2] = cur_price2

context_len = 512  # コンテキスト長の設定
horizon_len = 1  # 予測する期間の長さの設定
#horizon_len = 128  # 予測する期間の長さの設定

# TimesFMモデルの初期化と読み込み
tfm = TimesFm(
    context_len = context_len,
    horizon_len = horizon_len,
    input_patch_len = 32,
    output_patch_len = 128,
    num_layers = 20,
    model_dims = 1280,
    backend = bk_str,
)
tfm.load_from_checkpoint(repo_id="google/timesfm-1.0-200m")

# データの前処理
data_train = data_train[ohlc].dropna()  #欠損値を除去

if data_train.empty:
    raise ValueError("データが空です。期間を変更して再度試してください。")
data_train.loc[date_str1] = cur_price1
data_train.loc[date_str2] = cur_price2
#print("data length", len(data_train))
print( data_train )

if len(data_train) < context_len:
    raise ValueError(f"データの長さがコンテキスト長（{context_len}）より短いです。")

frequency_input = [0]  # データの頻度を設定（0は高頻度のデータ）

pred_len = 60
prev_pf = 23000
concat_pf = np.array([[]])
context_start = datetime.datetime(year = 2024, month = date_m, day = date_d )

for i in range(pred_len):
  tail = context_len + i
  context_begin = context_start - datetime.timedelta( days = i )
  context_end = context_start - datetime.timedelta( days = tail )
  context_data = data_train.loc[ context_end : context_begin ]

  # データの準備
  forecast_input = [context_data.values]

  # 予測の実行
  point_forecast, experimental_quantile_forecast = tfm.forecast(
    forecast_input,
    freq = frequency_input,
  )
  if i > 5 :
    if abs( 100 * ( point_forecast- prev_pf ) / prev_pf ) > discard_th:
      concat_pf = np.append( concat_pf, point_forecast )
      prev_pf = point_forecast
  else :
      concat_pf = np.append( concat_pf, point_forecast )
      prev_pf = point_forecast

concat_pf = np.flip( concat_pf )

#print("point_forcast", point_forecast[0], point_forecast[0].shape)
print("concat_pf", concat_pf, concat_pf.shape)

# 予測結果の表示
pred_len = concat_pf.shape[0]
forecast_dates = pd.date_range( start = data_train.index[ - pred_len ] + pd.Timedelta( days = shift_d ), periods = pred_len, freq='B' )
forecast_series = pd.Series(concat_pf, index=forecast_dates)
#print("forecast_series", forecast_series, forecast_series.index, forecast_series.values)
print("forecast_series", forecast_series)

plt.figure(figsize=(14, 7))
#plt.plot(data_train.index, data_train.values, label="Actual Prices")
plt.plot(data_all.index, data_all.values, label="Actual Prices")
plt.plot(forecast_series.index, forecast_series.values, label="Forecasted Prices")
start_date = datetime.date(2024, 4, 15)
end_date = datetime.date(2024, 11, 15)
plt.xlim(start_date, end_date)
plt.ylim(22000, 27000)
plt.xlabel("Date")
plt.ylabel("Price")
plt.legend()
plt.show()

[zhongtianda]

I did some modifications to predict next day price only. I run the inference every day instead of predicting multiple days ahead. So far it's accuracy is 67% (14/21) for up or down predictions. I find it very satisfactory. At least my investment performance increased a lot. You may need to adjust parameters in the code (shift_d, discard_th) every day for better accuracy.

Following code works in Google Colab

! git clone https://github.com/google-research/timesfm.git
%cd timesfm
!pip install -e .
!pip install utilsforecast
!pip install transformers accelerate bitsandbytes
!huggingface-cli login --token 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
#!pip install --upgrade numpy
!pip install numpy

import datetime
import yfinance as yf
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import timesfm
from timesfm import TimesFm

# 調整が必要なパラメータ(シフト量と抑制閾値はほぼ毎回調整が必要です)
# フラット抑制を高めると精度は上がるが、horizon(予測期間)が短くなる
date_m = 7 # month
date_d = 11 # day
shift_d = 7 # 時間軸のシフト量
discard_th = 0.00 # フラット抑制閾値(この変動率以下の予測値は切り捨て)
codelist = ["2558.T"] # 銘柄コード
ohlc = "Adj Close" # 4値

# 現在値(手動設定, 必要があれば)
date_str1 = "2024-7-10"
date_str2 = "2024-7-11"
prices1 = [24240.0, 24285.0, 24225.0, 25870.0]
prices2 = [24430.0, 24500.0, 24420.0, 26130.0]
bk_str = "gpu" # Backend

ohlc2i = {"Open": 0, "High": 1, "Low": 2, "Adj Close": 3}
cur_price1 = prices1[ohlc2i[ohlc]]
cur_price2 = prices2[ohlc2i[ohlc]]

start = datetime.date(2020, 1, 1)
end_t = datetime.date(2024, date_m, date_d )
#end_t = datetime.date.today()

data_train = yf.download(codelist, start = start, end = end_t)
data_all = yf.download(codelist, start = start, end = end_t)
print( data_train )

data_all = data_all[ohlc].dropna()  #欠損値を除去

if data_all.empty:
    raise ValueError("データが空です。期間を変更して再度試してください。")
data_all.loc[date_str1] = cur_price1
data_all.loc[date_str2] = cur_price2

context_len = 512  # コンテキスト長の設定
horizon_len = 1  # 予測する期間の長さの設定
#horizon_len = 128  # 予測する期間の長さの設定

# TimesFMモデルの初期化と読み込み
tfm = TimesFm(
    context_len = context_len,
    horizon_len = horizon_len,
    input_patch_len = 32,
    output_patch_len = 128,
    num_layers = 20,
    model_dims = 1280,
    backend = bk_str,
)
tfm.load_from_checkpoint(repo_id="google/timesfm-1.0-200m")

# データの前処理
data_train = data_train[ohlc].dropna()  #欠損値を除去

if data_train.empty:
    raise ValueError("データが空です。期間を変更して再度試してください。")
data_train.loc[date_str1] = cur_price1
data_train.loc[date_str2] = cur_price2
#print("data length", len(data_train))
print( data_train )

if len(data_train) < context_len:
    raise ValueError(f"データの長さがコンテキスト長（{context_len}）より短いです。")

frequency_input = [0]  # データの頻度を設定（0は高頻度のデータ）

pred_len = 60
prev_pf = 23000
concat_pf = np.array([[]])
context_start = datetime.datetime(year = 2024, month = date_m, day = date_d )

for i in range(pred_len):
  tail = context_len + i
  context_begin = context_start - datetime.timedelta( days = i )
  context_end = context_start - datetime.timedelta( days = tail )
  context_data = data_train.loc[ context_end : context_begin ]

  # データの準備
  forecast_input = [context_data.values]

  # 予測の実行
  point_forecast, experimental_quantile_forecast = tfm.forecast(
    forecast_input,
    freq = frequency_input,
  )
  if i > 5 :
    if abs( 100 * ( point_forecast- prev_pf ) / prev_pf ) > discard_th:
      concat_pf = np.append( concat_pf, point_forecast )
      prev_pf = point_forecast
  else :
      concat_pf = np.append( concat_pf, point_forecast )
      prev_pf = point_forecast

concat_pf = np.flip( concat_pf )

#print("point_forcast", point_forecast[0], point_forecast[0].shape)
print("concat_pf", concat_pf, concat_pf.shape)

# 予測結果の表示
pred_len = concat_pf.shape[0]
forecast_dates = pd.date_range( start = data_train.index[ - pred_len ] + pd.Timedelta( days = shift_d ), periods = pred_len, freq='B' )
forecast_series = pd.Series(concat_pf, index=forecast_dates)
#print("forecast_series", forecast_series, forecast_series.index, forecast_series.values)
print("forecast_series", forecast_series)

plt.figure(figsize=(14, 7))
#plt.plot(data_train.index, data_train.values, label="Actual Prices")
plt.plot(data_all.index, data_all.values, label="Actual Prices")
plt.plot(forecast_series.index, forecast_series.values, label="Forecasted Prices")
start_date = datetime.date(2024, 4, 15)
end_date = datetime.date(2024, 11, 15)
plt.xlim(start_date, end_date)
plt.ylim(22000, 27000)
plt.xlabel("Date")
plt.ylabel("Price")
plt.legend()
plt.show()

该代码存在问题，日期没对上，很多情况没考虑到。比如一开始预测的结果不可能和真实值差这么多，所谓的 67%准确率只是随机的结果罢了， 预测股票还是会有其它深度学习模型常用的滞后性。后面我再试试协变量的效果