Comprehensive analysis of DATFID's performance in the M5 Forecasting Accuracy challenge.
Report Date: August 2025
DATFID achieved exceptional performance in the M5 Forecasting Accuracy challenge, scoring 0.64 and ranking in the top 5% of all participants. This achievement was accomplished using zero-shot learning with full interpretability, demonstrating the power of transparent AI in real-world forecasting scenarios.
The M5 Forecasting Accuracy challenge is one of the most prestigious competitions in the field of time series forecasting. It uses real retail data from Walmart, comprising 3,049 products across 3 US states over 1,941 days, including sales, prices, promotions, and events. This represents a complex, multi-dimensional forecasting problem that mirrors real-world business scenarios.
Unlike traditional black-box machine learning models, DATFID employed an interpretable approach that provides clear mathematical relationships between variables. This transparency allows users to understand not just what the forecast predicts, but why it makes those predictions.
DATFID's approach leverages advanced panel data analysis techniques combined with interpretable AI algorithms. The system automatically identifies complex relationships between sales, pricing, promotions, and external factors, providing mathematical formulas that explain the forecasting logic.
y ~ α1*Intercept + β1*CPI + β2*Cultural + β3*National + β4*snap_CA + β5*snap_TX + β6*snap_WI + β7*wday + β8*month
Variable | Estimate | Standard Error | T statistic | P value |
---|---|---|---|---|
Intercept | -24,001.07 | 46.32 | 518.20 | 0.00 |
Variable | Estimate | Standard Error | T statistic | P value |
---|---|---|---|---|
CPI | 25,306.67 | 560.22 | 45.17 | 0.00 |
Cultural | -208.88 | 88.73 | 2.35 | 0.02 |
National | -756.18 | 78.36 | 9.65 | 0.00 |
snap_CA | 211.92 | 29.96 | 7.07 | 0.00 |
snap_TX | 256.12 | 32.91 | 7.78 | 0.00 |
snap_WI | 244.12 | 32.98 | 7.40 | 0.00 |
wday | -242.28 | 6.27 | 38.64 | 0.00 |
month | -16.18 | 3.66 | 4.42 | 0.00 |
CPI (25,306.67): Consumer Price Index has the strongest positive impact. A 1-unit increase in CPI leads to approximately 25,307 additional sales units, indicating strong price sensitivity.
Cultural (-208.88): Cultural events have a moderate negative effect, reducing sales by about 209 units when cultural events occur.
National (-756.18): National events have a stronger negative impact, reducing sales by about 756 units during national events.
snap_CA (211.92): SNAP benefits in California increase sales by about 212 units, showing positive impact of government assistance programs.
snap_TX (256.12): SNAP benefits in Texas have the strongest positive impact among states, increasing sales by about 256 units.
snap_WI (244.12): SNAP benefits in Wisconsin also show strong positive impact, increasing sales by about 244 units.
wday (-242.28): Day of the week has a negative effect, with sales decreasing by about 242 units as the week progresses.
month (-16.18): Month has a slight negative trend, with sales decreasing by about 16 units per month.
The M5 challenge results demonstrate DATFID's capability to handle enterprise-scale forecasting problems with both accuracy and transparency. This validation on real retail data proves that interpretable AI can compete with and outperform traditional black-box approaches while providing the clarity that business leaders need for decision-making.
DATFID's performance in the M5 challenge validates our approach to interpretable AI. By achieving top-tier accuracy while maintaining full transparency, we've demonstrated that businesses don't need to choose between accuracy and interpretability—they can have both.
For more information about DATFID's M5 performance or to discuss how our interpretable AI can benefit your business, please contact us:
Email: info@datfid.com
Address: Munich, Germany