Trained model on NFLX stock data from 2014-2021. Predictions are scaled wrongly.

[Rajmehta123]

Downloaded NFLX data from 2014 - 2021. Trained the model using default parameters. And the forecasting is wrongly scaled.

The NFLX stock value is around ~$515 on July 22, 2021, but the prediction value is around ~$20. Attached is the below graph for test and pred.

Below is the values for test and pred.

Pred: tensor([20.5919, 20.5882, 20.6023, 20.5925, 20.5604, 20.5682, 20.5184, 20.4771, 20.4323, 20.3718, 20.3013, 20.2565, 20.2335, 20.1562, 20.0954, 20.0564])

Test: tensor([492.4100, 498.3400, 500.7700, 497.0000, 508.8200, 512.7400, 518.0600, 527.0700, 533.0300, 533.5000, 528.2100, 533.5400, 533.9800, 541.6400, 535.9600, 530.7600])

[JulesBelveze]

Hey @Rajmehta123 thanks for pointing that out. There's indeed something weird going on with the scaling. I'm on holidays atm but will check that out as soon as I get back 🙂

PS: if you find out what is actually wrong please let me know. And if you have time to open PR that would fix it I would really appreciate it.

[Rajmehta123]

The dataset loader is using ColumnarTransformer in a pipeline. There is no way to inverse_transform it back. Also, I removed the normalization method just to see how the predictions are, but in vain. Even with no normalization, the predictions are in the range of $5-$20.

[JulesBelveze]

Sounds like a bug yeah. I'll check that out. Do you have a link where I can get the dataset you're using?

[JulesBelveze]

Also @Rajmehta123 could you provide me the configuration you're using please

[Rajmehta123]

Sure. I am attaching the link to the dataset.

https://drive.google.com/file/d/1gyEvwkyUKzWDUHOxCQcZXivoOv0sOmtz/view?usp=sharing

Below are the configs.

parser = argparse.ArgumentParser()
parser.add_argument("--batch-size", default=config["batch_size"], type=int, help="batch size")
parser.add_argument("--output-size", default=config["output_size"], type=int,
                    help="size of the ouput: default value to 1 for forecasting")
parser.add_argument("--label-col", default=config["label_col"], type=str, help="name of the target column")
parser.add_argument("--input-att", default=config["input_att"], type=lambda x: (str(x).lower() == "true"),
                    help="whether or not activate the input attention mechanism")
parser.add_argument("--temporal-att", default=config["temporal_att"], type=lambda x: (str(x).lower() == "true"),
                    help="whether or not activate the temporal attention mechanism")
parser.add_argument("--seq-len", default=config["seq_len"], type=int, help="window length to use for forecasting")
parser.add_argument("--hidden-size-encoder", default=config["hidden_size_encoder"], type=int,
                    help="size of the encoder's hidden states")
parser.add_argument("--hidden-size-decoder", default=config["hidden_size_decoder"], type=int,
                    help="size of the decoder's hidden states")
parser.add_argument("--reg-factor1", default=config["reg_factor1"], type=float,
                    help="contribution factor of the L1 regularization if using a sparse autoencoder")
parser.add_argument("--reg-factor2", default=config["reg_factor2"], type=float,
                    help="contribution factor of the L2 regularization if using a sparse autoencoder")
parser.add_argument("--reg1", default=config["reg1"], type=lambda x: (str(x).lower() == "true"),
                    help="activate/deactivate L1 regularization")
parser.add_argument("--reg2", default=config["reg2"], type=lambda x: (str(x).lower() == "true"),
                    help="activate/deactivate L2 regularization")
parser.add_argument("--denoising", default=config["denoising"], type=lambda x: (str(x).lower() == "true"),
                    help="whether or not to use a denoising autoencoder")
parser.add_argument("--do-train", default=True, type=lambda x: (str(x).lower() == "true"),
                    help="whether or not to train the model")
parser.add_argument("--do-eval", default=True, type=lambda x: (str(x).lower() == "true"),
                    help="whether or not evaluating the mode")
parser.add_argument("--data-path", default='nflx.csv', help="path to data file")
parser.add_argument("--output-dir", default=config["output_dir"], help="name of folder to output files")
parser.add_argument("--ckpt", default=None, help="checkpoint path for evaluation")

df = pd.read_csv(config["data_path"])
df = df.set_index('Date_Time')

if not os.path.exists(config["output_dir"]):
    os.makedirs(config["output_dir"])

ts = TimeSeriesDataset(
    data=df,
    categorical_cols=config["categorical_cols"],
    target_col=config["label_col"],
    seq_length=config["seq_len"],
    prediction_window=config["prediction_window"]
)

For the config file.:

device=torch.device('cuda' if torch.cuda.is_available() else 'cpu'),

categorical_cols=["day"],  # name of columns containing categorical variables
label_col=["Close"],  # name of target column
index_col="Date",

output_size=1,  # for forecasting

num_epochs=100,
batch_size=16,
lr=1e-5,
reg1=True,
reg2=False,
reg_factor1=1e-4,
reg_factor2=1e-4,
seq_len=10,  # previous timestamps to use
prediction_window=1,  # number of timestamps to forecast
hidden_size_encoder=128,
hidden_size_decoder=128,
input_att=True,
temporal_att=True,
denoising=False,
directions=1,

max_grad_norm=0.1,
gradient_accumulation_steps=1,
logging_steps=100,
lrs_step_size=5000,

output_dir="output",
save_steps=5000,
eval_during_training=True

[JulesBelveze]

Hey @Rajmehta123 there's a way to invert the scaler by directly accessing it:

self.preprocessor.named_transformers_["scaler"].inverse_transform(data)

I'm working on a "clean" way to convert the normalized values to the actualy ones 😃

[JulesBelveze]

@Rajmehta123 I've made the changes and managed to have it to work properly on my local with your configuration 😃 If you could check that it works properly for you before I merge the PR that would be awesome: https://github.com/JulesBelveze/time-series-autoencoder/pull/19

[JulesBelveze]

@Rajmehta123 feel free to reopen it if it still doesn't work for you.

[Rajmehta123]

Hey Jules. I just tried the PR but still, I am not able to produce the predictions. Can you share your config used on that NFLX dataset?

***** EVAL REPORT **** MSE = 194079.140625 residuals = -430.42496 loss = tensor(8.7637)

[Rajmehta123]

With the correct hyperparam, I could get a reasonable-performing model. All good now.