Forecasting AI in Forecasting

Lawrence Vanston, Ph.D. President, Technology Futures, Inc. [email protected]

International Symposium on Forecasting 2019

June 13-16, 2019 Thessaloniki, Greece

84 Waller St • Austin, Texas 78702 +1 512-415-5965 • www.tfi.com Copyright © 2019, Technology Futures, Inc. 1 Outline • Interest Areas • Technology Forecasting • Forecasting AI Performance • Forecasting Technology Adoption • Drivers and Constraints • Forecasting AI Adoption • Recommendations

Copyright © 2019, Technology Futures, Inc. 2 Interest Areas • Technology Forecasting • Communications • AI & New Wave Technologies Performance  Adoption  Impacts  Implications • AI in Forecasting • AI & Forecasting for Good • Fisterra Projects Art  Dance  Science  Technology

Copyright © 2019, Technology Futures, Inc. 3 Copyright © 2019, Technology Futures, Inc. 4 Forecasting Typology • Predictive Analytics • Demand Forecasting • Time Series Forecasting • Technology Forecasting / Long-Range Forecasting

Copyright © 2019, Technology Futures, Inc. 6 Technology Future’s Technology Forecasting Approach • Drivers and constraints • Adoption and substitution curves • Performance trends • Analogies • Expert opinion • Scenarios and ideation tools

Copyright © 2019, Technology Futures, Inc. 8 Forecasting AI Performance • Performance Trends in Computation (Exponential) • Performance Trends in AI (Typically Linear) • Problem Complexity and Performance (Reason Why)

• Linear progress is typical • Deep learning ANNs can cause discontinuities • Wide range of metrics to measure performance • Wide range of progress needed to reach human performance • More work to be done

1. Will the technology be adopted? 2. How big is the potential market? 3. When will it be commercially available? 4. How fast will it penetrate the market?

100% 90% 80% 70% 1.? 60% 50% 2. 40% 4. 30% Market Penetration 20% 10% 3. 0% 1995 2000 2005 2010 2015 2020 Time

Lawrence K. Vanston, Curt Rogers, and Ray L. Hodges, Advanced Video Services—Analysis and Forecasts For Terrestrial Service Providers, Technology Futures, Inc., 1995, p. 106. This graphic appears in Introduction to Technology Market Forecasting, 1996, p.25. Copyright © 2019, Technology Futures, Inc. 15 Ultra-HD Households (aka 4K)

100% 90% 80% 70% 60% 50% 40% 30% HDTV Ultra-HDTV Households Households HDTV 2017 20%

Percentage Households Percentage of 10% 0% 2000 2005 2010 2015 2020 2025 2030

Historical data sources: 2001-2004 Misc, Year 2005-2015 Leichtman Research Source: Technology Futures, Inc.

UHDTV Data Source (Red Squares): Strategy Analytics Copyright © 2019, Technology Futures, Inc. 16 Drivers and Constraints • What are the drivers for adoption? – How strong are they? • What are the constraints on adoption? – How strong are they? Can they be overcome? • What is the balance of drivers and constraints? – Will this change? • What are the important areas of uncertainty that need to be resolved? – How can these be addressed to everyone’s satisfaction? Copyright © 2019, Technology Futures, Inc. 17 www.telenor.com/en/innovation/ research/publications/telektronikk/ volume/telektronikk-3-4-2008 www.tfi.com

• Over 20 years old • Competitive in some applications with traditional methods • Improvement over time • Mixed AI/Statistical model won the 2018 M4 Competition (Slawek Smyl, Uber Technologies) • Presentations by major AI players at ISF 2018 in Boulder

• • • • • •

Copyright © 2019, Technology Futures, Inc. 22 Constraints on AI in Forecasting • Computation Intensive • Large data requirements for training • Over-fitting and instability • Not always the most accurate • Black Box • AI’s lack of insight • Not as familiar to forecasters Can these constraints be overcome?

Copyright © 2019, Technology Futures, Inc. 23 Constraints on AI in Forecasting • Computation Intensive  • Large data requirements for training  • Over-fitting and instability  • Not always the most accurate  • Black Box  • AI’s lack of insight  • Not as familiar to forecasters  Can these constraints be overcome? Current Assessment: Yes

Copyright © 2019, Technology Futures, Inc. 25 Recommendations • More forecasting on performance, adoption, impacts, and implications of AI and related technologies • More forecasting of AI for forecasting. • Make AI an essential part of forecasting learning, including Python. • Expand application of AR/VR and other technologies to forecasting • Re-expand forecasting to include less-than- statistical methods

Artificial Intelligence, especially machine learning, has been applied in forecasting for many years. Recent progress has been significant and raises the question of AI’s role in the future of forecasting. This paper provides a platform for addressing that question. We take a future-oriented view, examining the drivers and constraints for the continued adoption of AI. We also discuss how further progress in AI might or might not change the balance. Finally, we discuss the question of whether AI could substantially substitute for traditional forecasting methods, and, if so, what that means for the forecasting profession.