Neural networks can capture nonlinearity and complex seasonality that trip up simple baselines. Discover the architectures and tricks that turn raw history into reliable demand forecasts.
Compared with ARIMA, an LSTM can naturally learn ______ relationships over long lags.
static categorical
monotone decreasing cost
nonlinear
linear only
A practical trick for intermittent demand is to model zeros and positives separately, often called ______.
beam search
max pooling
two‑stage or zero‑inflated modeling
label smoothing
To capture day‑of‑week and holiday effects in a net, you can use ______ encodings.
SHA‑1 hashes of dates
grayscale image filters
learned embeddings for calendar features
random one‑hot noise
Temporal Convolutional Networks differ from standard CNNs by using ______ kernels.
2D image
causal (no peeking into the future)
non‑causal bidirectional by default
complex‑valued FFT only
When series counts are huge, a global model learns across items and a ______ layer adapts per item.
manual spreadsheet
QR code
GAN discriminator
small item‑specific head or embedding
For multi‑horizon forecasting, the seq2seq approach outputs ______.
calendar ICS files
raw gradients
a vector of future steps in one forward pass
only the next step
A common stabilization for multiplicative seasonality before feeding a net is a ______ transform.
z‑score then square
log
min‑max inverse only
square
Dropout during training primarily helps with ______.
calendar generation
regularization to reduce overfitting
batching across GPUs
data encryption
Teacher forcing in sequence models means ______ during training.
using only predicted values
freezing all embeddings
doubling the learning rate
feeding the true previous value into the decoder
To evaluate retail demand nets reliably, you should report scale‑free metrics like ______.
BLEU score
raw RMSE only
sMAPE or MASE
accuracy on shuffled data
Starter
You know the basics of sequence nets—keep practicing with real calendars.
Solid
Solid modeling instincts; refine embeddings and loss choices.
Expert!
Expert! You can design and ship neural forecasts with confidence.
