20  Homework 5: Deep Learning for Time Series

Due Week 9. Covers the feedforward and convolutional models, and the window-length experiment.

Your numbers will not match mine, and that is fine

Every result in this course depends on a random seed, and often also on your library versions, your hardware, and how many CPU threads your machine uses. Two people running identical, correct code will get different numbers — sometimes noticeably different.

What is graded is that you used the right model and the right procedure, and that your interpretation follows from the numbers you actually obtained. Quote your own output.

What is not graded is matching any particular value. Do not tune your way toward a number you saw on a slide or heard from a classmate — that is the one way to turn this into a wrong answer.

One caveat worth knowing. The broad ordering of models usually survives this variation. If yours comes out backwards from what the lecture found, say so in a sentence and offer a reason: either you have found something real about your run, or there is a bug worth hunting. Both are worth more marks than quietly reporting a number you do not believe.

20.1 Overview

This homework builds four neural network architectures for time series forecasting, progressing from simple to complex:

  1. Feedforward Network (FFN) — treats lagged sales as a flat feature vector
  2. 1D Convolutional Network (CNN) — slides a learned filter across the sequence to detect local temporal patterns
  3. LSTM — processes the sequence step-by-step, maintaining a memory of past states
  4. Transformer — attends to all time steps simultaneously, learning which past weeks are most informative

All four are implemented in PyTorch using the same sliding-window dataset infrastructure, making it straightforward to compare them on identical train/test splits.

Learning objectives:

  • Build a PyTorch Dataset and DataLoader for sliding-window time series data
  • Train a feedforward network on windowed input and read its loss curve
  • Implement a 1D convolutional network, and connect kernel size and dilation to its receptive field
  • Compare the FFN and the CNN on holdout RMSE, and say what the receptive-field experiment showed about window length

20.2 Dataset and Setup

We construct fixed-length sliding windows over each store-category series. Each input is a window of past weeks; the target is the next week’s units. Part A (FFN, CNN) uses a 26-week window; Part B (RNN, LSTM, Transformer) uses 52 weeks, so the sequence models get a full annual cycle.

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import math
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from sklearn.metrics import mean_squared_error
import warnings
warnings.filterwarnings('ignore')

from pathlib import Path

def data_path(name):
    for base in (Path.cwd(), *Path.cwd().parents):
        p = base / "data" / "processed" / name
        if p.exists():
            return p
    raise FileNotFoundError(
        f"Could not find data/processed/{name}. "
        "Run `python scripts/prep_m5.py` from the repository root first."
    )

RANDOM_STATE = 42
torch.manual_seed(RANDOM_STATE)
np.random.seed(RANDOM_STATE)

weekly = (pd.read_csv(data_path("m5_weekly.csv"), parse_dates=["week_start"])
            .sort_values(["store_id", "cat_id", "week_start"]))
weekly["series"] = weekly.store_id + "_" + weekly.cat_id

# The four channels every window carries through time
CHANNELS = ["units", "snap_days", "event_days", "avg_price"]

CUTOFF = pd.Timestamp("2015-05-17")     # same split as Homework 3 and 4

print(f"{weekly.series.nunique()} series | "
      f"{weekly.week_start.min().date()} to {weekly.week_start.max().date()} | "
      f"{len(weekly):,} rows")

Building the windows. For each series, slide a window of length W along the history: the window is the input, the units in week W+1 is the target. A window belongs to the training set if its target week falls before the cutoff.

Two mistakes that will not raise an error

Do not shuffle. DataLoader(..., shuffle=False) for every model here. Shuffling destroys the ordering the sequence models exist to exploit, and it will not warn you.

Standardize using training statistics only. Compute the mean and standard deviation on the training windows, then apply them to the test windows. Computing them over the full dataset leaks the test period into training — the same error as scaling outside a CV fold in Homework 4.

20.3 Initial Prompt for Your AI Assistant

Any assistant works — ChatGPT, Claude, Gemini, or Copilot, in an editor or in a browser tab. Paste the block below as your first message, before you ask anything else.

Initial Prompt for Your AI Assistant

Paste the following as your first message to your assistant. Save it as INITIAL_PROMPT.md in your submission folder.

I am working on Homework 5 Part 1 for ECON 8310: Business Forecasting at the
University of Nebraska at Omaha.

Dataset: M5/Walmart — weekly aggregated, 30 store-category series, at
`data/processed/m5_weekly.csv`. Columns: store_id, cat_id, state_id,
week_start, units (target), avg_price, snap_days, event_days, and lagged
units. I build sliding windows over each series: 26 weeks for Part A and
52 weeks for Part B, with four channels (units, snap_days, event_days,
avg_price). Temporal split at 2015-05-17, matching Homework 3 and 4.

This homework covers four neural network architectures for time series
forecasting: FFN, 1D CNN, LSTM, and Transformer encoder (all in PyTorch).
I will ask you questions one at a time. For each:
1. Write clean, well-commented PyTorch code that runs end-to-end
2. Explain what the output means so that I understand it — but do NOT draft the written
   answers I submit. The interpretation and business-recommendation sections are mine to
   write, in my own words.

My prompt budget for Part 1 is 9 (this context prompt is excluded). I will label
each question before asking.


Repository rules — these apply to every prompt in this thread:
- If you cannot write files or run git — a browser chat cannot — then the file and commit
  rules below are mine to carry out, not yours. Tell me what to write and what to commit,
  and I will do it. Everything else in this list still applies to you.
- My submission folder is `submissions/hw05_part1_<my_name>/`. Put everything there.
- Keep a `PROMPT_LOG.md` in that folder. For every prompt I send, append a block:
      ### Prompt <n> — YYYY-MM-DD HH:MM
      <my prompt, verbatim>
  Numbered from 1, in order. Never edit or delete an earlier entry. If your tool cannot write
  files, I will paste them in myself.
- After finishing the work for each prompt, make exactly one git commit:
  `hw05_part1 prompt <prompt_id>: <short summary>`
- If a prompt would take me past my budget of 9, do not do the work. Reply exactly
  PROMPT_LIMIT_REACHED, and still log that prompt.
- Never modify `AGENTS.md`, anything in `assignments/`, or anything in `policy/`.

Confirm by responding with: HW05P1_CONTEXT_ACK

20.4 Part A: FFN and CNN (window = 26 weeks)

How the code cells work

A cell that is empty except for a comment is where you write. The comment states the task; your code goes underneath it.

Some cells already contain setup code — a train/test split, a rename, a cutoff date — and end with a comment marking where your part begins. Leave the setup as it is and write below the comment. It is there so that everyone’s results are computed on the same basis.

Every cell runs as the document stands, so you can render at any point, finished or not. Doing that early and often is how you avoid finding a problem at the deadline.

The Your response blocks are written answers, not code. They are yours to write, in your own words — your assistant explains output to you, it does not draft these.

20.4.1 Q1 — Windows, Dataset and DataLoader (Suggested prompts: 3)

Implement a SeriesWindowDataset that turns the panel into supervised windows.

  • For each of the 30 series, slide a window of length W along the history
  • x: the window, shape (W, 4) — the four channels in CHANNELS
  • y: units in the following week
  • A window is training if its target week falls before CUTOFF

Standardize each channel using training statistics only, and standardize y the same way. Build DataLoaders with batch_size=64 and shuffle=False for both.

WINDOW_A = 26

class SeriesWindowDataset(Dataset):
    def __init__(self, df, channels, window, stats=None):
        # TODO: build windows per series; return (x, y) as float32 tensors
        pass

    def __len__(self):
        pass

    def __getitem__(self, idx):
        pass

# Instantiate train and test datasets + dataloaders

Your response: How many training windows do you get? It is fewer than the number of store-category-week rows before the cutoff — work out exactly how many you lose and explain where they go. (Hint: what has to exist before a window can start?) Then: why shuffle=False here, when almost every PyTorch tutorial you will find sets it to True?

20.4.2 Q2 — Feedforward Network (Suggested prompts: 2)

Flatten the window to a vector of length W × 4 and feed it to a plain network: Linear(W*4 → 128), ReLU, Dropout(0.2), Linear(128 → 64), ReLU, Dropout(0.2), Linear(64 → 1).

Train 30 epochs, Adam lr=1e-3, weight_decay=1e-4, MSE loss. Plot the loss curve and report test RMSE in original units (invert the standardization before computing RMSE).

class FFNForecaster(nn.Module):
    def __init__(self, window, n_channels):
        super().__init__()
        # TODO

    def forward(self, x):        # x: (batch, window, channels)
        # TODO
        pass

# Training loop (30 epochs)

Your response: Flattening throws away the fact that the window is ordered — the model cannot tell week 1 from week 26 except by position in a vector. Given that, how does its RMSE compare to the XGBoost result you recorded in Homework 4, which had the same disadvantage?

20.4.3 Q3 — 1D Convolutional Network (Suggested prompts: 2)

Build a CNNForecaster: Conv1d(4, 32, kernel_size=5), ReLU, Conv1d(32, 64, kernel_size=3), ReLU, then flatten and a final Linear to one output.

Note the input shape: Conv1d expects (batch, channels, time), so you must transpose the (batch, time, channels) tensors from your DataLoader.

Train as in Q2 and report test RMSE.

class CNNForecaster(nn.Module):
    def __init__(self, n_channels):
        super().__init__()
        # TODO

    def forward(self, x):        # remember the transpose
        # TODO
        pass

Your response: Lecture 8 measured roughly a 500-RMSE penalty for replacing Flatten with AdaptiveAvgPool1d(1). Explain in one or two sentences why pooling costs accuracy here, in terms of what the pooling operation discards.

20.4.4 Q4 — Receptive Field and Window Length (Suggested prompts: 2)

Retrain the CNN at windows of 13, 26 and 52 weeks. Report test RMSE for each.

# Retrain CNN for each window length; collect RMSE

Your response: Does a longer window always help? Your two convolution layers see about seven weeks of context regardless of how long the window is — explain how that limits what a longer window can buy, and what you would change architecturally to use a 52-week window properly.

20.5 Submission Checklist

Not graded, but worth doing: commit locally once per prompt, hw05_part1 prompt <id>: <summary>. Nobody sees these — they are your own record of which prompt produced what, and they make it far easier to retrace your steps when something breaks.