Published: July 11, 2026

Forecast Models Used in Market Research

Every market research report that lands on an executive's desk carries a number attached to a future year a market size, a growth rate, a demand curve and almost nobody asks how that number was actually produced. Behind it sits a specific forecasting model, chosen from a toolkit that has expanded dramatically over the past few years. Time series methods such as ARIMA and exponential smoothing still dominate stable, data-rich categories, but gradient-boosted trees, transformer-based architectures, and now pre-trained foundation models are increasingly doing the heavy lifting on volatile or multivariate problems. The stakes of picking correctly are not academic. The 2022 M5 forecasting competition, still the most cited benchmark in the field, found that carefully tuned classical statistical methods routinely matched or beat far more complex deep learning models on univariate retail series, a result that keeps surprising analysts who assume newer always means better. Understanding which model does what and where each one breaks down is the difference between a market forecast that holds up and one that quietly falls apart six months after publication.

Time series models: the workhorse of stable markets

Time series forecasting remains the default starting point whenever at least two years of consistent historical data exist, and for good reason: these models are fast to train, mathematically transparent, and don't require external variables to function. ARIMA Auto-Regressive Integrated Moving Average combines autoregression, differencing for stationarity, and moving-average error correction into a single framework, with its seasonal variant SARIMA adding explicit periodic terms and SARIMAX layering in outside covariates such as price or promotions. Holt-Winters exponential smoothing is the other classical staple, prized for adapting flexibly to shifting trend and seasonal patterns without heavy parameter tuning. A comparative study of retail-trade time series across seven European countries found that these traditional linear approaches ARIMA extensions, classical decomposition with multiple linear regression, and Holt-Winters remain highly competitive against nonlinear alternatives, concluding that claims of neural networks' automatic superiority in retail forecasting were exaggerated. In applied settings the accuracy gains can be substantial regardless of method: one industrial demand-forecasting deployment using time series regression reached over 90 percent accuracy on product-specific sales, while a trade-promotion forecasting rollout for a consumer goods company hit 92 percent accuracy and lifted service levels to 98.6 percent.

Econometric and associative regression models

Where time series models look inward at a single variable's own history, econometric and associative models look outward at the relationships between variables interest rates, employment levels, marketing spend, competitor pricing to explain and project market behavior. Linear regression remains the baseline associative tool, useful for quantifying something as concrete as a five-dollar sales lift for every one-dollar increase in ad spend, while logistic regression handles binary outcomes such as purchase versus non-purchase and is widely used in marketing to model conversion probability. A comparison of four models built to forecast stock market returns ARIMA, polynomial regression, decision tree regression, and random forest regression found that random forest delivered the best combination of accuracy and resistance to overfitting, while polynomial regression tended to overshoot and decision trees fit almost too well to historical noise to generalize forward. Econometric models are especially favored for high-level, long-horizon planning, such as a real estate developer weighing a residential project against forecasted interest rates and inflation, because the coefficients themselves carry economic meaning that pure statistical curve-fitting does not provide.

Qualitative and judgmental methods still anchor new-market forecasts

Quantitative models assume there is enough historical data to model in the first place, and in market research that assumption frequently fails. For genuinely new categories a first-of-its-kind product, an unproven technology, or entry into an unfamiliar geography analysts fall back on qualitative approaches: expert panels, the Delphi method, sales force composite estimates, jury-of-executive-opinion, and life-cycle analogy forecasting. Delphi modeling in particular is the standard choice for innovation-heavy sectors such as health sciences, AI, and renewable energy, where structured, anonymous rounds of expert estimation converge on a number that no historical dataset could have produced. Total Addressable Market sizing, the figure investors scrutinize most closely, is itself built on three parallel qualitative-quantitative hybrids — top-down analysis starting from broad industry reports, bottom-up calculation from customer counts and pricing, and value-theory estimation based on the economic value delivered to the buyer and rigorous practice now calls for cross-referencing estimates across firms such as Gartner, Forrester, and IDC, since convergence across independent sources is treated as the real signal of confidence, not any single number in isolation.

Machine learning and the foundation-model shift

The most significant change in forecasting methodology since 2022 has been the move from bespoke statistical models toward machine learning approaches that treat forecasting as a general prediction problem rather than a time series-specific one. Gradient-boosted decision tree frameworks LightGBM, XGBoost, and CatBoost convert historical data into lag features and rolling statistics, then apply standard supervised learning, an approach that frequently outperforms both classical statistics and deep learning on tabular, feature-rich forecasting problems. Deep learning architectures including N-BEATS, Temporal Fusion Transformers, and DeepAR pushed accuracy further on complex multivariate series, with the Temporal Fusion Transformer's Variable Selection Networks offering a rare combination of strong performance and interpretability that lets analysts see exactly which input variables are driving a given prediction. The newest layer is foundation models: large architectures pre-trained across billions of time points from many different domains that can forecast an unfamiliar series with no task-specific retraining, mirroring how large language models generalize across text tasks. This shift is playing out inside the research industry itself one 2026 survey of insight professionals found that 62 percent report most or some of their team already using AI tools, 64 percent say their AI tool count grew over the past year, and 46 percent expect their AI budgets to increase again in 2026, even as researchers describe the mood as cautious optimism rather than blind enthusiasm.

Choosing the right model — and why none of them are universal

No single forecasting model wins across every market research scenario, and the research literature is unusually consistent on this point: different models capture different aspects of a series, and the choice should follow the data, not the fashion. Time series models need at least two years of stable historical data; econometric models suit long-range planning where causal drivers matter more than pattern-matching; qualitative methods are the only realistic option for markets with no track record at all; and machine learning approaches earn their complexity premium mainly when the problem is genuinely multivariate or feature-rich. Model evaluation itself follows standard error metrics MAPE, MAE, RMSE, and MASE with lower values indicating tighter fit, and best practice always benchmarks a chosen model against a naive baseline, such as simply carrying the last known value forward, to prove the added complexity is actually earning its keep. Given how much capital allocation, product timing, and investment decisions ultimately trace back to a forecasted number, the discipline of matching model to market and being honest about the resulting margin of error remains one of the most underrated skills in applied market research.

Frequently Asked Questions

What is the most commonly used forecasting model in market research?

There is no single default, but time series models such as ARIMA and exponential smoothing remain the most widely used starting point whenever at least two years of consistent historical data is available, because they are fast, interpretable, and require no external variables. Econometric and associative regression models are the next most common choice for markets where outside economic drivers matter.

Are machine learning models more accurate than traditional statistical models?

Not automatically. The 2022 M5 forecasting competition and multiple follow-up studies found that well-tuned classical methods like ARIMA and Holt-Winters frequently match or beat deep learning models on univariate and low-dimensional series. Machine learning and foundation models tend to pull ahead specifically on complex, multivariate, feature-rich problems rather than simple single-variable forecasts.

How do analysts forecast a market that has no historical data?

They rely on qualitative and judgmental methods rather than quantitative models. Common approaches include the Delphi method, expert panels, sales force composite estimates, and life-cycle analogy forecasting, along with Total Addressable Market sizing built from top-down, bottom-up, and value-theory approaches, often cross-checked across multiple independent research firms.

What is the difference between time series and econometric forecasting?

Time series models predict a variable using only its own historical pattern trend, seasonality, and past error terms. Econometric and associative models instead forecast using the relationships between the target variable and outside economic or business factors, such as interest rates, marketing spend, or employment data, which makes them better suited to explaining why a market will move, not just modeling that it will.

What accuracy metrics are used to evaluate a forecasting model?

The standard metrics are MAPE (Mean Absolute Percentage Error), MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and MASE (Mean Absolute Scaled Error), where lower values indicate better fit. Best practice also compares any chosen model against a naive baseline, such as simply repeating the last known value, to confirm the added modeling complexity is actually adding value.

What are foundation models in forecasting, and how are they different from traditional models?

Foundation models are large architectures pre-trained on billions of time points across many different industries and data types, allowing them to generate forecasts for a new, unfamiliar series without task-specific retraining. This mirrors how large language models generalize across text tasks, in contrast to traditional models like ARIMA or Prophet, which must be built and tuned individually for each specific dataset.

Is AI adoption actually widespread in the market research industry yet?

Adoption is accelerating quickly but not yet universal. A 2026 industry survey of insight professionals found 62 percent report that most or some of their team already uses AI tools, 64 percent saw their number of AI tools increase over the past year, and 46 percent expect their AI budget to grow further in 2026, with researchers describing the overall mood as cautious optimism rather than full-scale replacement of traditional methods.