The method

Why the rate wins

Every GetFocus verdict rests on one number: the Technology Improvement Rate. This page shows where it comes from, why it forecasts winners, and how far it's been validated. Explainable maths on public patent data — not a generative-AI guess.

Methodology co-developed with MIT — 30 years of peer-reviewed research.

Watch · The method, explainedJard van Ingen · Co-founder & CEO
01 · The foundation

Built on three decades of MIT research

01

The question they asked

Can you measure, objectively, how fast a technology improves? MIT built real cost-and-performance histories for 28 technology domains — decades of data each, assembled by hand over six years.

02

What they found

Every technology improves exponentially, each at its own remarkably stable rate. A Moore's law for everything, not just chips.

03

Why it lets us forecast

Because the rates are stable, they tell you where a technology is heading long before the market makes it obvious. Head-to-head, the faster improver historically took over — every time.

“Frankly, I didn't expect to be so precise.”

Gordon Moore — on how exactly Moore's law held, decade after decade
Peer-reviewed foundation
Benson & Magee — MIT, 2010–2017
Measuring and estimating technology improvement rates from real cost-and-performance data. Several papers.
Triulzi, Alstott & Magee — 2020
Predicting those rates from patent centrality metrics — the leap from measurement to forecast.
GetFocus — 2020 onward
Automated the measurement, refined the method beyond the papers, and validated it on 150+ domains.
02 · The signal

Performance is the snapshot. The rate is the signal.

Judge by current performance and the incumbent looks safe for years. Judge by the rate and the crossover is inevitable from day one.

“IT'LL NEVER CATCH UP” “OOPS” YEARS → PERFORMANCE →
Incumbent · +5%/yrChallenger · +50%/yr
IMPROVEMENT RATE, %/YR — MIT-MEASURED 01 · TELECOM 02 · STORAGE 03 · LIGHTING Optical fibre 62 Copper telecom 14 Flash storage 42 Magnetic storage 32 LED lighting 34 Incandescent 4 0 62 %/YR
Won its categoryThe slower rival it displaced

The rates span ~20×. Optical telecom improved 62%/yr — doubling every ~17 months. Milling machines: 3%/yr — doubling every ~23 years. Where two technologies do the same job, the quicker improver historically took over. Every time.

03 · The data

Patents record progress years before the market shows it

You can only measure improvement rates empirically when it's already too late. The leading indicator hides in the patent record.

01

Early

Before a technology reaches the market, it's in the patents — and almost everything is patented.

02

Complete

The vast majority of technical progress finds its way into the patent literature, across every field and country.

03

Standardised

Patents are one of the few globally consistent records of invention — structured, standardized and citation-linked.

The paradox: counted one by one, patents lag the market — the breakthroughs show up late. But the undercurrent of disruption forms in the citation network years before those breakthrough patents are ever filed. That undercurrent is what we read.

04 · The dataset

An improvement rate is only as good as the set of patents behind it

Before we calculate an improvement rate, we need to know which patent families belong to the technology in question, and which only look like they do. Get this wrong and every number downstream is wrong. To do this well, we need world-class patent search and filtering. Here's how that works at GetFocus.

Stage 01 · Recall

AI search by meaning rather than keywords

The technology is written down as a concept instead of as a keyword string. Every patent family in the corpus sits in a semantic space; we pull the ones related to that concept in meaning and cast an overly wide net on purpose.

Real concept query · sodium-ion for automotive

“Sodium-ion battery cells for electric vehicle traction, covering layered transition-metal oxide, polyanionic and Prussian blue analogue cathode materials, hard carbon anodes, sodium-salt electrolytes and additives, and cell and electrode engineering that raises energy density, fast-charging capability, cycle life and low-temperature performance of sodium-ion cells toward the requirements of low-cost and standard-range passenger electric vehicles.”

VECTOR SPACE · ONE DOT = ONE PATENT FAMILY YOUR CONCEPT FURTHER IN MEANING
Candidates retrievedLeft in the corpus

Recall is deliberately generous at this stage. A family that never gets retrieved can never be judged, so the net is cast wider than the final scope.

Stage 02 · Precision

Large language models read every patent in full, then make the final relevance call

A language model reads every retrieved family in full: title, abstract, claims and description. It then holds it against one written instruction, family by family. Nothing is skimmed and nothing is sampled. A Boolean string can only match the words an inventor happened to choose, while a model that has read everything can tell a sodium-ion traction cell from a grid-storage cell easily.

Real filter instruction · excerpt

“A patent qualifies if it claims sodium-ion battery technology, sodium-based cathodes, hard carbon or other sodium-storage anodes, sodium electrolytes, or cell and manufacturing innovations specific to sodium-ion cells: AND the described cell is plausible for vehicle traction. […] Exclude filings whose stated application is exclusively stationary, grid-scale or long-duration storage, backup power or consumer electronics. Exclude sodium-sulfur and sodium-metal-halide (ZEBRA) chemistry. Exclude lithium-ion patents that mention sodium only as a dopant. Return: INCLUDE or EXCLUDE, and nothing else.”

Instruction × patent textThe call
WO-2024230880-A1
Pre-sodiated anode and composite cathode active material· sodium cathode and anode chemistry
INCLUDE
DE-102023112154-A1
Pre-sodiation and de-sodiation for reduced formation loss· manufacturing specific to sodium-ion cells
INCLUDE
US-8986885-B2
Lithium ion battery· lithium-ion chemistry, not sodium-ion
EXCLUDE
US-20240145687-A1
Composite oxide for a lithium-ion secondary battery, electronic device and power storage system· lithium-ion, stated uses are electronics and stationary storage
EXCLUDE
US-20240356070-A1
Lithium ion battery· lithium-ion electrolyte and cathode, no sodium-storage claim
EXCLUDE

Each call is made on that one family's own text against the same instruction, not a blanket query, and never on the abstract alone.

What you end up with

One dataset per technology, complete and clean

Complete because recall is cast wide and nothing is sampled. Clean because every family in it was read and ruled in deliberately.

Every relevant family, in any language. Retrieved by meaning across the global corpus, not by the words one inventor happened to use.
Nothing in it that does not belong. Adjacent-domain filings, and filings that merely mention the technology in passing, were read and excluded.
One row per invention, not per office. The same invention filed in fifteen countries is one patent family, so a company that files widely does not read as a company that invents more.
Rates you can actually compare. Every dataset is drawn the same way, so an improvement-rate gap between two routes is a real difference in the technologies, not in the search.
05 · The mechanics

From 180M+ patents to one number per route

01→

Build the relevant patent set

You sign off on the list of candidate technologies. Our LLM agents read and filter global patent data to uncover the entire track record of every route.

02→

Read the citation network

We compute two signals from how those patents cite one another — cycle time and knowledge flow.

03

Estimate the rate

Those two metrics yield the improvement rate for any route — long before it's visible in the market.

Metric 01

Cycle time

The gap, in years, between an invention and the earlier one it improves on.

Shorter cycle time → advancing faster
Metric 02

Knowledge flow

How much later inventions build on it — forward citations across the network.

More knowledge flow → bigger leaps

We run the whole pipeline for you — the dashboard, the reports, the alerts. AI gathers and reads the patents; the maths does the forecasting. The forecast is an explainable, peer-reviewed MIT method — not a generative-AI guess.

06 · Validation

Measured far beyond the original papers

28
Domains MIT measured by hand — six years of work
150+
Domains GetFocus has empirically measured — 5× the original research
R² ≈ 0.7
Our patent-based forecast vs the measured rate

0.7 means the forecast explains ~70% of the variation in measured improvement rates — strong for a leading indicator, and far better than gut feel. And the method calls losers, not just winners: the rotary engine drew ~$50M of GM's money in 1970, but its rate flagged a dead end from the start.

07 · Proof from history

Built only on data available at the time

Every patent carries a filing date, so the record can be replayed. In every case, filing volume — the consensus — pointed at the wrong route. The rate pointed at the one that became the standard.

Powertrain
Hydrogen vs lithium
The whole industry was filing into hydrogen — the consensus. Automakers committed billions to fuel cells.
Lithium improved ~2× faster the entire time — a decade ahead of Tesla and BYD.
Read the case →
Imaging
Film vs digital — the Kodak case
Film kept out-filing digital until its sales had already collapsed. The consensus held long after it was wrong.
Digital's rate was higher from the very start — flagged 20 years before disruption.
Read the case →
Storage
HDD vs SSD
HDDs out-filed SSDs until around 2010 — the consensus stuck with magnetic storage deep into the disruption.
SSDs improved ~2× faster from the start — the rate would have flagged Western Digital's 2009 flash move years earlier.
Read the case →
Engines
Rotary vs piston — calling a loser
Rotary drew real money — GM paid ~$50M in 1970 (~$400M+ today) for Wankel licences; Mazda chased it for decades.
Its improvement rate never came close to the piston engine — a dead end, flagged from the start.
Read the case →
Read all seven cases →
08 · Due diligence

The questions your team will ask

Is this another generative-AI black box?+
No. AI agents do the reading — gathering and filtering 180M+ patents. The forecast itself is an explainable regression on two citation metrics, built on peer-reviewed MIT research and validated on 150+ domains at R² ≈ 0.7. Every number traces back to the patent record.
Patent counts lag the market. Why would this lead it?+
Counted one by one, patents do lag — that's the consensus signal, and it's usually wrong. We don't count patents; we read the citation network. Cycle time and knowledge flow form years before breakthrough filings appear. In every backtest, filing volume pointed at the losing route while the rate pointed at the winner.
Can it call losers, or only winners?+
Both. The rotary engine drew serious investment for decades — GM licences, Mazda's persistence — but its improvement rate never came close to the piston engine. The rate flagged a dead end from the start.
What do you need from us?+
One strategic question, in plain language — “Which battery chemistry wins for low-cost EVs?” You sign off on the route list our agents scout. That's the entire input. We run the pipeline, stand up the dashboard, and keep it alive.
How fast is the first radar?+
Weeks, not the months a one-off study takes. Then it never goes stale: monthly signal summaries, a quarterly executive review with our analysts, and trigger alerts the moment a threshold you set is crossed.
What about fields where patenting is thin?+
The vast majority of technical progress reaches the patent literature, across fields and countries. Where coverage is genuinely thin for a specific route, we flag it during scoping — before it can distort a verdict, not after.
Don't take our word for it

Give us your toughest technology decision. See what comes back in 48 hours.

One live decision, one decision-grade brief: the radar, the improvement rates, the triggers. Built on your domain, not a demo dataset.

Get a free 48-hour Decision Brief

Free · 48 hours · preliminary scoping — validated radars come with a subscription