Can't get TiDE to handle future covariates

[StephenRogers1]

I have been trying on multiple datasets to get TiDE to handle covariates as stated but I can't seem to enable this. I am using the TiDE code. In particular, I produced a dataset that simply produces the target by adding two covariates x,y and cannot recover the target, this is similar to a realistic dataset where it performs worse with covariates despite the covariates having a strong relationship with the target.

Code to reproduce:

np.random.seed(42)
x = np.random.normal(0, 1, 13308)
y = np.random.normal(0, 1, 13308)
date = pd.date_range(start='2020-01-01', periods=13308, freq='15min')
T = x + y
df = pd.DataFrame({'date': date, 'x': x, 'y': y, 'T': T})
df.to_csv('datasets/covariates.csv', index=False)

look_back=384
horizon=384
cov_cols='x, y'
target_cols='T'
ts_cols="${target_cols}, ${cov_cols}"
#!/bin/bash
python3 -m train \
    --train_epochs=0 \
    --ts_cols="${ts_cols}" \
    --num_cov_cols="${cov_cols}" \
    --transform=false \
    --layer_norm=true \
    --holiday=false \
    --dropout_rate=0.0 \
    --batch_size=512 \
    --hidden_size=512 \
    --num_layers=1 \
    --hist_len=$look_back \
    --pred_len=$horizon \
    --dataset=covariates \
    --decoder_output_dim=8 \
    --final_decoder_hidden=16 \
    --num_split=1 \
    --learning_rate=0.00003012706619800982 \
    --min_num_epochs=-1 \
    --normalize=false

In Train.py DATA_DICT

    "covariates": {
        "boundaries": [10646, 11977, 13308],
        "data_path": "./datasets/covariates.csv",
        "freq": "15min",
        "num_cov_cols": ["x", "y"],
        "ts_cols": ["T"],
        }

As you can see I added to both the DATA_DICT entry and the shell command, but neither allowed use of the covariate. The error is higher (1.3 MAE) than without covariates (1.0 MAE).

I want to use TiDE with covariates as its performance seems good, however, it seems they are either specified some other way or there is some other issue.

Thanks,
Stephen