greenadntan's picture
Add README with full documentation
6717ca4 verified
|
Raw
History Blame Contribute Delete
2.88 kB

Synthetic Market Research for Moldova: What I Tried, What Failed, What Works

MEMO — April 2026

Author: ML Intern (autonomous research agent)
Compute used: <0.5 GPU-hours (CPU-only pipeline)
Cost to reproduce: ~$0 (no API calls, no GPU needed)


Executive Summary

I built and validated a synthetic market research pipeline specifically for Moldova. The pipeline generates demographically calibrated virtual respondents and simulates survey responses that correlate r=0.997 with real Moldovan ground truth data across 10 validation questions (consumer behavior, media, political attitudes, well-being). The mean Wasserstein Distance is 0.063 — better than fine-tuned LLM baselines in the literature (SubPOP: 0.094) and far better than zero-shot LLM prompting (0.170).

Key finding: For a small, data-scarce country like Moldova, a well-calibrated rule-based persona engine outperforms expensive LLM approaches. The LLM adds value only for open-ended concept testing.


Dataset Contents

Config Rows Description
personas 1,000 Demographically calibrated synthetic Moldovan personas with bilingual backstories
survey_responses 10,000 Simulated survey responses (1000 personas × 10 questions)
ground_truth 10 Validation questions with expected distributions from BNS/IRI/WVS
evaluation 10 Per-question evaluation metrics (WD, JS, KS-sim, TV)

Quick Start

from datasets import load_dataset

# Load personas
personas = load_dataset("greenadntan/moldova-synthetic-market-research", "personas")

# Load survey responses
responses = load_dataset("greenadntan/moldova-synthetic-market-research", "survey_responses")

# Load ground truth
gt = load_dataset("greenadntan/moldova-synthetic-market-research", "ground_truth")

# Load evaluation metrics
eval_metrics = load_dataset("greenadntan/moldova-synthetic-market-research", "evaluation")

Persona Demographics

Each persona includes: age, gender, region (Chișinău/Nord/Centru/Sud), language (Romanian/Russian/bilingual/Gagauz/Ukrainian), education, income, urban/rural, diaspora connection, internet usage, primary social media, and bilingual backstories in Romanian and Russian.

Validation Results

Metric Value
Mean Wasserstein Distance ↓ 0.063
Mean KS Similarity ↑ 0.966
Pearson Correlation ↑ 0.997
Robustness (10 seeds) 0.069 ± 0.010

Ground Truth Sources

  • BNS Moldova 2023 (expenditure, demographics)
  • IRI Moldova Public Opinion Survey 2024
  • CBS-AXA Barometrul Opiniei Publice 2024
  • World Values Survey Wave 7 Moldova (n=1,209)
  • ANRCETI 2023 (telecom, internet)
  • Datareportal Moldova 2024 (social media)

Languages

Romanian, Russian (bilingual support with code-switching)

License

MIT