Forecasting and Trend Analysis Questions

Covers the full set of forecasting methods and trend analysis techniques used to project future business performance and validate predictions. Topics include time series techniques, moving averages, exponential smoothing, regression analysis, seasonality adjustments, straight line and percentage growth projections, and regression to the mean. Also covers driver based and bottom up build up models, top down allocations, scenario and sensitivity modeling, judgement based forecasts, and sales forecasting approaches such as pipeline based forecasting, deal probability weighting, collaborative forecasting, and historical trending. Includes model validation and accuracy assessment techniques, methods to identify and correct forecast bias, back testing against historical data, assessing confidence intervals and uncertainty, and communicating assumptions and forecast limitations.

MediumTechnical

78 practiced

Describe methods to detect forecast bias across products or regions. Once bias is detected (forecasts consistently overestimating by 7%), outline 4 concrete statistical or process steps you would take to correct or mitigate that bias in production forecasts.

Sample Answer

Detecting bias- Compute error metrics by product/region: MFE (mean forecast error), MAPE, MASE, and track signed errors (y_hat - y). MFE = (1/N) Σ(y_hat - y) shows direction.- Visualize heatmap/time-series of MFE by product × region to spot systematic patterns.- Run statistical tests: t-test on mean error ≠ 0 per segment; control charts (CUSUM or EWMA) to detect persistent shifts.- Fit a simple regression of residuals on features (product, region, seasonality) to quantify drivers of bias.

If forecasts are consistently overestimating by ~7%, four concrete corrective steps

1) Apply a calibrated multiplicative correction (quick fix)- Compute bias factor b = mean(y_hat)/mean(y) ≈ 1.07. Adjust future forecasts: y_hat_adj = y_hat / b (or y_hat * (1 - 0.07)).- Validate on holdout period and monitor MFE reduction.

2) Retrain model with bias-aware loss or sample weighting- Add a penalty for systematic overforecast (asymmetric loss) or reweight training examples from biased segments so model learns ground truth distribution.- Example: use weighted MSE where weights are higher for under-predicted bins.

3) Introduce hierarchical/Bayesian model or per-segment intercepts- Fit a hierarchical model with product/region random effects or a parametric bias term per segment: residual_ij = alpha_j + eps; subtract estimated alpha_j from forecasts.- This captures persistent offsets while sharing strength across sparse segments.

4) Productionize monitoring + feedback loop- Deploy automated monitoring (daily/weekly): track MFE, MAPE by segment, alert if |MFE|>threshold.- Implement A/B test: serve adjusted forecasts to a subset of regions/products, measure downstream KPIs (inventory turns, stockouts) to ensure business benefit.- Schedule periodic recalibration (e.g., rolling 8–12 week recalculation) and add new features if bias correlates with omitted covariates (promotions, lead-time changes).

These steps combine immediate mitigation, model improvement, per-segment calibration, and ops controls to both fix the 7% overforecast and prevent recurrence.

EasyTechnical

110 practiced

You have a CRM pipeline table: (deal_id, amount, stage, probability, expected_close_date). Describe and implement (in pseudocode) an algorithm to compute a weighted sales forecast by month from this pipeline using deal probabilities and historical stage conversion rates. Describe assumptions and corrections for known salesperson optimism bias.

Sample Answer

Approach summary:- Use historical data to compute stage-to-close conversion rates and salesperson calibration factors.- For each open deal, compute an adjusted win probability = reported_probability * stage_conversion_rate(stage) * salesperson_calibration.- Optionally apply aging/time-to-close decay and cap probabilities.- Multiply by deal amount and sum by forecast month (expected_close_date).

Pseudocode (Python-like):

python

# inputs: pipeline_rows: list of dicts {deal_id, amount, stage, probability, expected_close_date, owner}
# history: past deals with final outcome (won/lost) and recorded stage history
# params: min_prob, max_prob, aging_half_life_months

# 1) compute stage conversion rates from history
stage_conv = compute_stage_conversion_rates(history)  # dict stage -> P(close | reached stage)

# 2) compute salesperson calibration: observed_win_rate / avg_reported_prob
owner_calib = compute_owner_calibration(history)  # dict owner -> factor (<=1 reduces optimism)

# 3) helper: aging correction based on days until expected close
def aging_factor(days):
    return 1 - 0.5 ** (days / (30 * aging_half_life_months))

# 4) produce weighted forecast by month
forecast = defaultdict(float)
today = now()
for d in pipeline_rows:
    base_p = d['probability'] or global_avg_reported_prob
    stage_p = stage_conv.get(d['stage'], global_stage_avg)
    owner_p = owner_calib.get(d['owner'], 1.0)
    days = (d['expected_close_date'] - today).days
    af = aging_factor(days) if days > 0 else 1.0
    adj_p = base_p * stage_p * owner_p * af
    adj_p = min(max(adj_p, min_prob), max_prob)
    month = d['expected_close_date'].strftime('%Y-%m')
    forecast[month] += d['amount'] * adj_p

Assumptions & corrections:- Stage conversion rates are stationary enough to use recent N months (use weighted recency).- Salesperson optimism: calibrate per-owner factors = observed_win_rate / avg_reported_prob (regularize with Bayesian shrinkage to avoid overfitting).- Probabilities reported by reps reflect subjective likelihood; combine with objective stage conversion to reduce bias.- Apply aging: deals far in future get dampened; stale deals (no activity) get reduced probability.- Validate by back-testing (MAPE) and adjust caps, half-life, and recency windows. Edge cases: missing owner/date/probability handled with defaults and conservative assumptions.

HardTechnical

66 practiced

Design and describe an LSTM (TensorFlow/Keras) model architecture for multi-horizon forecasting of sales with exogenous features (promotions, price, holiday indicators). Include input shapes, loss functions for multi-step outputs, handling missing values, and training/validation strategies appropriate for time-series data.

Sample Answer

To forecast multi-horizon sales with exogenous features using an LSTM in TensorFlow/Keras, use a seq-to-seq pattern: encoder processes historical windows, decoder (or a dense head) predicts H future steps. Key points: input shapes, masking for missing values, multi-step loss (optionally horizon-weighted), and time-series-aware validation (rolling-origin).

Approach:1) Inputs:- past_sales: shape (batch, T_in, 1)- exog_past: shape (batch, T_in, E_past) (promotions, price history, holidays)- exog_future: shape (batch, H, E_future) (known future exog like planned promotions/holidays)Concatenate past features into (batch, T_in, 1+E_past).

2) Model (simple encoder + repeat + decoder):

python

import tensorflow as tf
from tensorflow.keras import layers, Model

T_in = 28   # input window
H = 14      # forecast horizon
E_past = 2
E_future = 3

past_input = layers.Input(shape=(T_in, 1+E_past), name="past_input")
future_exog = layers.Input(shape=(H, E_future), name="future_exog")

# Masking to handle padded/missing timesteps (use NaN->mask_value or use mask tensor)
x = layers.Masking(mask_value=0.0)(past_input)
x = layers.LSTM(128, return_sequences=False)(x)
# optional: repeat context and concatenate future exog
context = layers.RepeatVector(H)(x)
decoder_input = layers.Concatenate(axis=-1)([context, future_exog])
y = layers.LSTM(64, return_sequences=True)(decoder_input)
out = layers.TimeDistributed(layers.Dense(1))(y)

model = Model([past_input, future_exog], out)

Loss for multi-step outputs:- Use timestep-wise MSE: loss = mean_over_horizons(MSE). To emphasize near-term accuracy, use a weighted MSE:

python

horizon_weights = tf.constant([...], dtype=tf.float32)  # shape (H,)
def weighted_mse(y_true, y_pred):
    se = tf.square(y_true - y_pred)  # (batch,H,1)
    se = tf.squeeze(se, -1) * horizon_weights  # broadcast weights
    return tf.reduce_mean(se)

- Handle missing target values with a mask: pass sample_weight or multiply loss by mask (1 for observed, 0 for missing).

Handling missing values:- For exogenous features: impute (forward-fill, median) and add binary indicators for imputed points.- For targets: if occasional NaNs, mask them in loss or remove affected windows. Use layers.Masking for input masks; for irregular time intervals consider time-lag feature.- Normalize numeric features per time series (scaler saved per series).

Training & validation strategies:- Use rolling-origin (walk-forward) validation: create multiple train/val splits by moving the forecast origin, preserving temporal order.- Avoid random shuffles; set shuffle=False in fit and use stateful batches carefully or stateless LSTM with sequence batching.- Use TimeSeriesGenerator or custom windowing to produce (past_window, future_exog, future_targets).- Use callbacks: EarlyStopping (monitor val_loss), ReduceLROnPlateau, and checkpointing.- Evaluation metrics: per-horizon RMSE/MASE and aggregated business metrics (e.g., bias on promo periods).- For cross-series learning (many SKUs), include SKU embeddings or categorical features.

Other considerations / trade-offs:- Stateful LSTM can capture long continuity but complicates batching; prefer stateless with longer windows.- Transformer/TCN may outperform LSTM for long horizons.- Calibration: quantify uncertainty via quantile loss (pinball) or Monte Carlo dropout / outputting prediction intervals.

This architecture balances simplicity and practical needs: masking/imputation for missingness, horizon-aware loss, and robust time-series validation. Adjust hyperparameters (layers, units, window length) and incorporate business constraints (shelf capacity, promo rules) as needed.

MediumTechnical

66 practiced

Describe three methods to compute prediction intervals for forecasts and discuss their assumptions and limitations: (1) analytical residual-based intervals, (2) bootstrap residuals, and (3) quantile regression or quantile forests. Which would you pick for heavy-tailed error distributions?

Sample Answer

Analytical residual-based intervals- Approach: Fit model, estimate residual variance (often assuming homoscedastic Gaussian errors) and use forecast ± z*·SE(forecast) (or model-specific formulas for ARIMA, linear regression).- Assumptions: Correct model form, independent identically distributed (i.i.d.) Gaussian residuals (or known distribution), correct variance estimate.- Limitations: Sensitive to model misspecification, heteroscedasticity, non-normal or heavy-tailed errors → intervals can be too narrow or miscalibrated.

Bootstrap residuals- Approach: Fit model, collect residuals, resample residuals (or block-bootstrap for dependent data), generate many synthetic future paths, compute empirical percentiles for prediction intervals.- Assumptions: Residuals are representative of future errors; for time series, dependence handled by block bootstrap or model-based bootstrap.- Limitations: Computationally heavier; if residuals are heteroscedastic or nonstationary, naive bootstrap fails; small samples give noisy percentiles; block choices affect bias/variance.

Quantile regression / quantile forests- Approach: Directly estimate conditional quantiles (e.g., 5th/95th) of target given features using quantile regression, gradient-boosted quantile trees, or quantile random forests.- Assumptions: Model flexible enough to capture conditional distribution; fewer distributional assumptions than analytical methods.- Limitations: Requires sufficient data across covariate space for stable quantile estimates; may underperform for extreme tails if rare events absent; interpretation depends on covariates included.

Recommendation for heavy-tailed errors- Prefer quantile-based methods (quantile regression or quantile forests) or bootstrap with tail-aware resampling. Quantile methods are robust because they estimate conditional quantiles directly without Gaussian assumptions; if using bootstrap, ensure resampling preserves tail behavior (use larger bootstrap runs, model-based heavy-tail residuals, or Pareto-tail modeling). Analytical intervals are least appropriate for heavy tails.

EasyTechnical

78 practiced

You need to choose a primary accuracy metric for monthly revenue forecasting across many products where some revenues are very small or zero. Compare MAE, RMSE, MAPE and MASE and recommend one or a combination, explaining advantages and pitfalls for skewed data and zeros.

Unlock Full Question Bank

Get access to hundreds of Forecasting and Trend Analysis interview questions and detailed answers.

Join thousands of developers preparing for their dream job.