TFT Beginner needs some insights

[ankur-connect]

Hi all,

I'm working on a large-scale time series forecasting problem using Temporal Fusion Transformers (TFTs), and I’d appreciate guidance on both the modeling philosophy and practical setup.

Philosophical Concern — Does the premise hold?
Forecasting future revenue using:

~2 years of past daily data per time series (~750 days)

Only future-known variable (macroeconomic conditions that too are the estimates)

Scale: ~100,000 time series

Concern:

Revenue (or demand) is often influenced by future events, not just past patterns

Is it philosophically sound to rely on past + very less future context for predicting something that’s highly forward-looking? Is there any EDA that I can do that somewhat addresses the feasibility of this forecasting problem?

Feature Engineering — Am I feeding junk?
Engineered ~250 features including:

Common stat features (mean, std, % change)

Custom types like TVKR, TVUR, SC, SR, etc.

All derived from a small set of primary sources

Questions:

How well does TFT handle large, highly correlated feature sets?

Should I be doing feature selection beforehand, or let the model learn? Or does this whole feature engineering make sense in case of TFT?

Cohorting / Clustering — Should I segment first?
Considering clustering the time series before modeling:

Based on seasonality, trend shape, or statistical profile

Idea: train one TFT per cluster

Questions:

Does clustering help reduce variance or improve convergence with TFTs?

Or is it better to let a global TFT learn from all series together?

Practical Setup — What are the right starting knobs?
Dataset size:

~100,000 time series × 750 timesteps = ~75 million rows

~250 features (past + known future + static)

Planning to use neuralforecast

Questions:

What’s the way to estimate optimal hyperparameters to start with given the data dimensions? For this scale, does using TFT make sense?