The prediction result of spike data is unstable

[nai-kon]

Thank you for this great project.

I have monthly data which is usually 0 and spikes in a constant cycle.
Here is a sample with 6 month cycle.

date, value
2020/4/1, 0
2020/5/1, 70
2020/6/1, 0
2020/7/1, 0
2020/8/1, 0
2020/9/1, 0
2020/10/1, 0
2020/11/1, 71
2020/12/1, 0
2021/1/1, 0
2021/2/1, 0
2021/3/1, 0
2021/4/1, 0
2021/5/1, 71
2021/6/1, 0

Here is the inference result. The peak is predicted correctly, but 0 is unstable.

I think this is because there is little (or no) spike data like this in the training data, as general time series data tends to constantly fluctuate.
Is there a way to make 0 stable?

sample data and code

test.csv

df = pd.read_csv("test.csv")
df.ds = pd.to_datetime(df.ds)
df["unique_id"] = "dummy"

train_data = df[df.ds < "2024-1"]

tfm = timesfm.TimesFm(
      hparams=timesfm.TimesFmHparams(
          backend="gpu",
          per_core_batch_size=32,
          horizon_len=15,
          num_layers=50,
          use_positional_embedding=False,
          context_len=2048,
      ),
      checkpoint=timesfm.TimesFmCheckpoint(
          huggingface_repo_id="google/timesfm-2.0-500m-pytorch"),
)

pred = tfm.forecast_on_df(
    inputs=train_data,
    freq="MS",
    value_name="y",
    num_jobs=-1,
)