Method

Why technology improvement is predictable, and how to measure it

GetFocus

·

May 2, 2025

·

9

min read·Updated

October 22, 2026

Why technology improvement is predictable, and how to measure it

Stay ahead of the curve

Get exclusive insights delivered directly to your inbox. Join thousands of tech leaders.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

TL;DR

  • How do you measure how fast a technology is improving? Pick one measure of performance per unit of cost, track it for at least a decade and fit an exponential curve. The slope is the annual improvement rate.
  • Is technology improvement predictable? Largely: MIT researchers found all 28 domains they studied improving exponentially, each at its own rate, from 3.1% a year for electric motors to 65.1% for optical telecommunications.
  • Can patents predict the rate? Yes, through citations: how quickly a domain's patents get cited, and how recent the patents they build on are, track its improvement rate closely. The number of patents does not.
  • How can you tell real improvement from hype? Hype shows up in news, funding and filing volume. Improvement shows up in performance per dollar. Compare rates, and read the noise as context.
  • What does this mean for an R&D decision? Compare the rates of the competing routes before you fund one. In GetFocus's lighting backtest, LED was improving at 32.5% a year in 1990 and incandescent at 8.6%.

How do you measure how fast a technology is improving? Track one measure of performance per unit of cost over time, plot it on a log scale and fit a straight line: the slope is the annual improvement rate. GetFocus estimates the same rate from patent citation data, with a method built on MIT research, so it works even for technologies too young to have reliable performance data.

That rate is the clearest signal of where a technology is heading. This guide covers what it measures, why it is predictable, how to calculate it two ways, and what it changes in an R&D decision.

What technology improvement means in numbers

Technology improvement is the rate at which a technology gets better at its job, measured as performance per unit of something scarce: cost, weight, volume or energy. Lighting is measured in lumens per dollar, solar modules in watts per dollar, data links in bits per second per dollar. A useful measure describes the job the technology does, so competing technologies can be compared on the same scale.

Two observations made this idea famous:

  • Moore's law. In 1965 Gordon Moore noted that the number of components on an integrated circuit was doubling at a steady pace. Transistor counts have doubled roughly every two years since.
  • Wright's law. In 1936 Theodore Wright found that the cost of building an aircraft fell by a constant percentage each time cumulative production doubled. Solar modules follow the same pattern: their price fell from $106 per watt in 1976 to $0.38 in 2019, about 20% cheaper with each doubling of installed capacity.

Both describe exponential improvement: a constant percentage gain per year, or per doubling of production. On a log scale, exponential improvement draws a straight line, and the slope of that line is the rate.

Is technology improvement predictable? What MIT's research found

Researchers at MIT have spent more than a decade measuring technology improvement, first across dozens of domains and then across the whole patent system. Four findings matter for anyone making a technology bet:

  1. Improvement is exponential, and every domain has its own rate. Magee and colleagues measured 28 technology domains over decades. Every one improved exponentially, at rates from 3.1% a year for electric motors to 65.1% for optical telecommunications.
  2. Rates differ enormously. Extending the method to 1,757 domains, Singh, Triulzi and Magee found rates from about 2% to 216% a year. More than 80% of domains improve at less than 25% a year, and the fastest are predominantly based on software and algorithms.
  3. A domain's rate holds steady. MIT's technology rates project describes the rate as "constant (at least to a good approximation) over time in a domain". A rate measured today is a reasonable guide to the next decade.
  4. Forecast errors are measurable too. Testing forecasts on 53 technologies, Farmer and Lafond found that errors follow a consistent distribution, so a forecast can come with realistic error bars.

The predictable part is the pace of a technical route: how fast its performance compounds. Which company wins inside that route is a separate question, and one that improvement rates leave open.

How to measure a technology's improvement rate, step by step

There are two ways to get the number. Use performance data when it exists. Use patents when it does not, or to check the first answer.

Method 1: from performance data

  1. Name the job and the measure. Write down the function and a measure of it per unit of cost, such as lumens per dollar.
  2. Collect at least ten years of data points. Product specifications, cost series and published records all work, as long as they measure the same thing.
  3. Plot the data on a log scale and fit a straight line. An exponential trend becomes a straight line; scatter around it is normal.
  4. Read off the annual rate. Fit a line to the logarithm of the measure against time; the annual rate is e^slope − 1. As a quick check, (end value ÷ start value)^(1 ÷ years) − 1 gives the average rate between two points. A measure that doubles every two years is improving at about 41% a year.

The common mistakes are mixing measures, picking only record-breaking data points and using too short a window.

Method 2: from patents and their citations

Young technologies rarely have ten years of clean performance data. Their patents tell you more, because every patent cites the earlier inventions it builds on, which turns the patent record into a map of how knowledge flows.

  1. Define the domain. Gather the patents that describe one technical approach to one function.
  2. Map the citation network. Link each patent to the earlier patents it cites, across the whole patent record.
  3. Measure the signals that track improvement. How quickly the domain's patents are cited by later patents, how recent the patents it builds on are, and how central it sits in the flow of knowledge. Benson and Magee found that citations within three years of publication tracked improvement rates closely (r = 0.76), as did the recency of cited patents (r = 0.72). Later MIT work found centrality in the citation network a highly reliable predictor.
  4. Convert the signals into an annual rate. Calibrate against domains where measured performance data exists.
  5. Compare against the competing routes. A rate means most next to the rates of the alternatives, and next to any performance data you have.

GetFocus runs this measurement for every route that could do the job you are deciding on. It reads two signals in the patent record, cycle time and knowledge flow, turns them into an improvement rate for each route and sets it next to each route's maturity. The full method is on our method page.

Worked example: LED against incandescent lighting

In GetFocus's lighting backtest, the data available in 1990 put LED lighting's improvement rate at 32.5% a year and incandescent lighting's at 8.6%. At 32.5% a year, performance doubles roughly every two and a half years. At 8.6%, it takes more than eight.

Improvement rate

What the 1990 improvement rates implied by 2007

Performance index, 1990 = 1, on a log scale, where a steady rate draws a straight line. Each line compounds at the annual rate GetFocus measured for 1990 in its lighting backtest.

LED, 32.5% a yearIncandescent, 8.6% a year
What the 1990 improvement rates implied by 2007 Performance index on a log scale from 1990 to 2007. LED compounding at 32.5% a year reaches about 120 times its 1990 level by 2007; incandescent at 8.6% a year reaches about 4 times. 1×10×100× 19901995200020052007 LED 120× Incandescent 4× 2007: GE bets on a better filament 1990 · LED 1.0× · Incandescent 1.0×1991 · LED 1.3× · Incandescent 1.1×1992 · LED 1.8× · Incandescent 1.2×1993 · LED 2.3× · Incandescent 1.3×1994 · LED 3.1× · Incandescent 1.4×1995 · LED 4.1× · Incandescent 1.5×1996 · LED 5.4× · Incandescent 1.6×1997 · LED 7.2× · Incandescent 1.8×1998 · LED 9.5× · Incandescent 1.9×1999 · LED 12.6× · Incandescent 2.1×2000 · LED 16.7× · Incandescent 2.3×2001 · LED 22.1× · Incandescent 2.5×2002 · LED 29.3× · Incandescent 2.7×2003 · LED 38.8× · Incandescent 2.9×2004 · LED 51.4× · Incandescent 3.2×2005 · LED 68.1× · Incandescent 3.4×2006 · LED 90.3× · Incandescent 3.7×2007 · LED 119.6× · Incandescent 4.1×
See the numbers
YearLEDIncandescent
19901.0×1.0×
19954.1×1.5×
200016.7×2.3×
200568.1×3.4×
2007119.6×4.1×
Source: GetFocus lighting backtest, getfocus.eu/proof/lighting. Constant-rate compounding, shown to illustrate what each rate implies.getfocus.eu

Compounded from 1990, LED's rate implied about 120 times the starting performance by 2007, against about 4 times for incandescent. In 2007, GE bet on a better filament. The lighting backtest shows the full analysis, and our proof page has backtests in other domains.

Why patents, and why citations rather than counts

Patents are an early, global and standardised record of invention. They are usually filed years before a product ships, they cover every major economy, and each one lists the prior art it improves on.

Counting them is the trap. In the same MIT study, the number of patents in a domain was not a significant predictor of its improvement rate (r = 0.33, p = 0.085), while the citation signals above were. Filing volume measures how much attention a field gets. Citation dynamics measure how fast it is moving. Our post on using patent trends to forecast emerging technologies goes deeper on reading patent signals.

How to tell if a technology is improving or just hyped

Most technology news measures attention. Use it as context, and look for the measurement underneath.

SignalWhat it measuresHow to read it
News coverage and search interestAttentionContext. Attention peaks often arrive before the performance does.
Funding rounds and startup countsInvestor appetiteContext. Capital follows stories as well as results.
Patent filing volumeActivityWeak on its own. Volume did not predict improvement in MIT's study.
Performance per dollar over timeImprovementThe signal, once ten years of data exist.
Patent citation dynamicsImprovement, earlierThe signal before performance data exists.

Three questions cut through most breakthrough claims. Which performance measure improved? At what annual rate, over how many years? And how does that rate compare with the technology it hopes to replace?

What predictable improvement means for R&D decisions

Maturity tells you where a technology is today; its technology readiness level (TRL) is the standard scale. The improvement rate tells you where it is going. Plot both, and four decisions fall out:

  • Fund routes that are mature and improving fast.
  • Monitor routes that are immature but improving fast, because they will not stay immature.
  • Phase out routes that are mature and improving slowly.
  • Ignore routes that are immature and improving slowly.

Our analysis of fourteen routes to low-carbon ethylene shows the logic on a real portfolio: electrochemical CO₂-to-CO improving at 43.6% a year against 13.0% for conventional steam cracking.

Rates also tell you when a crossover is likely. If a challenger trails the incumbent by a factor G and improves faster, the years to parity are roughly ln(G) ÷ ln((1 + challenger's rate) ÷ (1 + incumbent's rate)). A route ten times behind, improving at 30% a year against 10%, catches up in about 14 years. It is a rule of thumb that assumes steady rates, which is what the evidence above supports. For a fuller method of setting technologies side by side, see how to compare competing technologies objectively.

FAQ

How do you measure how fast a technology is improving?

Choose one measure of performance per unit of cost, such as lumens per dollar, collect at least ten years of data and fit an exponential trend. The slope of that trend on a log scale gives the annual improvement rate. For young technologies without that history, the rate can be estimated from patent citation data. GetFocus does this for every competing route in a domain.

How can I tell if a new technology is actually improving or just being hyped?

Separate attention from performance: news coverage, funding and patent filing volume measure attention and activity. Improvement shows up as a rising performance-per-cost measure, or, earlier, as fast citation dynamics in the patent record. Ask for the measure, the rate and the comparison with the incumbent's rate.

Is technology improvement predictable?

The rate of improvement is, to a good approximation. MIT research across 28 domains found every one improving exponentially at its own, roughly constant rate, and a later study extended rate estimates to 1,757 domains covering about 97% of US patents. Specific events, such as which company wins or when a product launches, are much harder to predict.

What is a technology improvement rate?

The annual percentage by which a technology's performance per unit of cost (or weight, volume or energy) improves. Rates run from about 2% to more than 200% a year depending on the domain; most domains improve at less than 25% a year.

Can patent data predict how fast a technology will improve?

Yes, through citation data. In Benson and Magee's 2015 study, citation-based measures predicted domains' improvement rates, with good predictive power for more than ten years ahead. The number of patents in a domain did not predict its rate.

How does an AI platform for patent analysis help in recognizing disruptive technologies?

Disruption shows up as a challenger that is still immature but improving faster than the incumbent. An AI platform can measure improvement rates for the incumbent and every challenger from patent citation data, and flag that pattern years before the challenger matches the incumbent on performance. GetFocus does this for every route in a domain; the LED and incandescent example above shows the pattern.

What is the difference between Moore's law and Wright's law?

Moore's law describes improvement over time: transistor counts doubling roughly every two years. Wright's law describes improvement with experience: cost falling by a fixed percentage each time cumulative production doubles. Both describe constant-percentage improvement, Moore's per unit of time and Wright's per doubling of output; when production grows steadily, they trace the same kind of curve.

Compare the rates before you commit

Bring us one technology decision you are weighing. Within 48 hours, a free Decision Brief maps the competing routes by improvement rate, so the choice rests on measured signals.

Sources

  1. Magee, C.L., Basnet, S., Funk, J.L. and Benson, C.L. (2016). Quantitative empirical trends in technical performance. Technological Forecasting and Social Change, 104, 237–246. doi.org/10.1016/j.techfore.2015.12.011
  2. Benson, C.L. and Magee, C.L. (2015). Quantitative determination of technological improvement from patent data. PLOS ONE, 10(4), e0121635. doi.org/10.1371/journal.pone.0121635
  3. MIT News (15 April 2015). Patents forecast technological change. news.mit.edu
  4. Triulzi, G., Alstott, J. and Magee, C.L. (2020). Estimating technology performance improvement rates by mining patent data. Technological Forecasting and Social Change, 158, 120100. doi.org/10.1016/j.techfore.2020.120100
  5. Singh, A., Triulzi, G. and Magee, C.L. (2021). Technological improvement rate predictions for all technologies: use of patent data and an extended domain description. Research Policy, 50(9), 104294. doi.org/10.1016/j.respol.2021.104294
  6. MIT technology rates project, overall methodology. technologyrates.mit.edu
  7. Farmer, J.D. and Lafond, F. (2016). How predictable is technological progress? Research Policy, 45(3), 647–665. doi.org/10.1016/j.respol.2015.11.001
  8. Moore, G.E. (1965). Cramming more components onto integrated circuits. Electronics, 38(8).
  9. Wright, T.P. (1936). Factors affecting the cost of airplanes. Journal of the Aeronautical Sciences, 3(4), 122–128.
  10. Roser, M. (2020). Why did renewables become so cheap so fast? Our World in Data. ourworldindata.org/cheap-renewables-growth
  11. MIT Schwarzman College of Computing (2 August 2021). A comprehensive study of technological change. computing.mit.edu
  12. Our World in Data. Transistors per microprocessor. ourworldindata.org
  13. GetFocus lighting backtest. getfocus.eu/proof/lighting

Ready to see it in action?

Bring us one technology decision you're weighing. We'll send back the competing routes mapped by improvement rate and maturity.

Talk to us