Why I'm Writing This
Time series forecasting sounds complicated and often it is but when you're building predictive pipelines for real users, clarity matters more than complexity. At PACX.AI, we've worked hard to make time-based predictions work in a no-code setup, and I wanted to share how we actually think about it under the hood.
This post walks through the core logic - nothing too abstract, just real building blocks you can follow. Whether you're building a forecasting engine yourself or evaluating a platform that promises it, this breakdown will help you understand what’s really going on.
⚠️ Note: This is a simplified version of our process. In production, PACX.AI includes additional layers like edge-case handling, outlier filtering, and robust feature scaling that aren’t shown here.
A Simple Dataset
Let’s assume you have a dataset like this:
Each entity_id (say, a user, customer, or machine) generates activity with some associated value over time.
✅ Step 1: Define Sampling Points
We start by picking the timeline we want to sample on weekly, monthly, daily, etc. Let's go with weekly sampling here.
We first find the minimum and maximum date in the dataset and generate sampling anchors dates that fall on a Monday between the min and max.
These will serve as the "checkpoints" for creating forecast rows.
✅ Step 2: Pair Entities with Eligible Dates
Now, not every entity existed at the beginning of the dataset. So we only allow an entity to be paired with sampled dates on or after their first activity.
If Entity B first appeared on 2024-01-05, it shouldn’t be sampled on 2024-01-01.
We join the sampled dates with each entity’s first known activity date, filtering out anything earlier
✅ Step 3: Compute Future Target (Next 7 Days)
Now for each (entity_id, sampled_date) pair, we want to compute the total value in the 7 days that follow.
For example:
For A on 2024-01-01, we look at activity from 2024-01-01 to 2024-01-07
For B on 2024-01-08, we look at activity until 2024-01-14
We sum up the value column in that range. That gives us the target — the number we want to predict.
Step 4: Generate Past Features (Last 30 Days)
To make a good prediction, we also need to know how the entity was behaving before the sampled date.
So we take a 30-day lookback window and calculate things like:
Average value in the last 30 days
Count of events
Maybe even last recorded value
If no past data is found (say it’s the first ever activity), we set it to NULL.
🧾 Final Output
After all this, our training dataset looks like:
This is your model-ready dataset. Each row is a prediction scenario: given the past (avg_30_day_value), can we predict the future (future_7_day_total)?
Why This Works:
- This kind of pipeline works well because:
- It respects entity-level history
- It makes time-aware splits without leakag
- It produces structured, explainable training data In our case, we’ve built a fully automated AI-powered no-code interface around this at PACX.AI, so the user can just uploads a CSV or ETL from any data source of his choice, confirms few things and everything else happens automatically. But under the hood, this is the logic we still follow.
Final Thoughts
Time series forecasting isn’t magic — it’s just structured logic applied carefully over time.
Whether you're building, evaluating, or just exploring, I hope this breakdown made things clearer and more practical.
If this helped you or you want to connect, reach out. Always open to good conversations.
— Sai Ram, Co-founder @ PACX.AI
🔗Connect on LinkedIn
Top comments (0)