Findings
Analysis
Four things 4,630 launch records say about the last 65 years of spaceflight, plus a topic model over mission names, and the limits of all of it.
Finding one
The launch rate collapsed in 1979 and stayed flat for thirty six years. The era that began in 2016 averages 111 launches a year, more than the 81 a year flown at the height of the space race.
Finding two
Spaceflight is far more concentrated than it looks. Ten organisations out of 62 flew 80 per cent of everything, and one of them, the Soviet strategic rocket forces, flew 38 per cent on its own.
Finding three
Flying more does not make you more reliable. It makes your reliability knowable. Volume buys precision, not performance.
Read down the chart and the pattern is not that busy operators are better. It is that busy operators are measured. RVSN USSR flew 1,777 times, so its 90.8 per cent is pinned to a range of 89.4 to 92.1, a spread of under three points. Starsem flew ten times without a failure, and its interval runs from 72.2 per cent to 100. The second operator has the better headline number and almost no evidence behind it.
Eleven operators in the full record have a perfect launch history. The largest of them, Blue Origin, has 21 flights, and even that only supports a claim of 84.5 per cent or better. Four of the eleven have flown exactly twice, where a perfect record is consistent with a true rate as low as 34 per cent.
The relationship can be tested rather than eyeballed. Ranking operators by launches against success rate and testing the correlation by permutation, 20,000 shuffles:
| Operators included | Count | Spearman rho | p |
|---|---|---|---|
| At least 1 launch | 62 | 0.2644 | 0.0369 |
| At least 5 launches | 41 | 0.5294 | 0.0004 |
| At least 10 launches | 29 | 0.1959 | 0.3064 |
Across all 62 operators there is a weak positive correlation. Among the 29 that have flown at least ten times it disappears: rho of 0.20 at p of 0.31, which is no evidence of a relationship at all. The apparent link is produced almost entirely by one-and-done operators whose single flight failed, not by veterans outperforming newcomers.
Two of the busiest operators in the record make the point directly. General Dynamics flew 251 times at 80.9 per cent and the US Air Force flew 161 times at 80.1 per cent, both far below the 89.9 per cent average, and both with more flights than SpaceX. Experience is not a substitute for the state of the engineering.
Finding four
Getting to orbit is now a routine engineering problem. Across 65 years, including the entire experimental era, nearly nine in ten attempts worked.
Operators in full
The ten busiest, and how often they worked
| Operator | Launches | Successes | Success rate |
|---|---|---|---|
| RVSN USSR | 1,777 | 1,614 | 90.8% |
| CASC | 338 | 318 | 94.1% |
| Arianespace | 293 | 282 | 96.2% |
| General Dynamics | 251 | 203 | 80.9% |
| VKS RF | 216 | 202 | 93.5% |
| NASA | 203 | 186 | 91.6% |
| SpaceX | 182 | 172 | 94.5% |
| US Air Force | 161 | 129 | 80.1% |
| ULA | 151 | 150 | 99.3% |
| Boeing | 136 | 131 | 96.3% |
All 62 operators, with confidence intervals, are in operator-reliability.csv.
Modelling
What an unsupervised model finds in a launch name
The modelling step asks a narrow question: if you hand a model nothing but the text of mission and rocket names, with no dates and no labels, does it recover the structure of the space age on its own?
Latent Dirichlet Allocation was run with gensim over two corpora, mission names and rocket names, at five topics each and 50 passes, after stopword removal and dropping tokens of three characters or fewer. It is a clustering method, not a predictor. It groups words that occur together and assigns each name a mixture over those groups.
It does recover the structure. The mission model separates a Cold War cluster (discoverer, intercosmos, apollo, surveyor) from a modern constellation cluster (starlink, oneweb, crew), and isolates the Soviet cosmos designation so strongly that it takes 77 per cent of the weight in its own topic. The rocket model splits along the same seam, Soviet workhorses in one topic and the Falcon and Ariane era in another.
Topics recovered
| Corpus | Topic | Highest weighted terms |
|---|---|---|
| mission | 0 | discoverer 0.061 · shijian 0.047 · intercosmos 0.038 · apollo 0.027 · satellite 0.024 · lunar 0.021 · surveyor 0.021 · pioneer 0.019 · oscar 0.018 · transit 0.016 |
| mission | 1 | starlink 0.097 · group 0.059 · soyuz 0.045 · intelsat 0.041 · oneweb 0.024 · gambit 0.016 · test 0.014 · explorer 0.011 · crew 0.011 · spacex 0.011 |
| mission | 2 | molniya 0.074 · nrol 0.048 · yaogan 0.043 · jilin 0.033 · flight 0.032 · rideshares 0.027 · eutelsat 0.019 · shiyan 0.017 · demo 0.017 · goes 0.012 |
| mission | 3 | gaofen 0.061 · beidou 0.054 · meteor 0.043 · progress 0.041 · gsat 0.029 · iridium 0.023 · chinasat 0.019 · jcsat 0.015 · galileo 0.013 · zenit 0.013 |
| mission | 4 | cosmos 0.771 · luna 0.017 · goniets 0.007 · gemini 0.006 · himawari 0.005 · quasar 0.004 · tiros 0.004 · gatv 0.003 · vostok 0.003 · proton 0.003 |
| rocket | 0 | atlas 0.452 · agena 0.094 · proton 0.084 · centaur 0.060 · briz 0.039 · iias 0.030 · rocket 0.026 · starship 0.022 · prototype 0.022 · launcherone 0.019 |
| rocket | 1 | cosmos 0.554 · tsyklon 0.230 · vostok 0.138 · scout 0.022 · able 0.008 · redstone 0.006 · mercury 0.005 · lambda 0.004 · black 0.002 · arrow 0.002 |
| rocket | 2 | block 0.266 · falcon 0.180 · ariane 0.105 · molniya 0.101 · electron 0.035 · pslv 0.032 · curie 0.030 · zenit 0.030 · shepard 0.021 · kuaizhou 0.020 |
| rocket | 3 | voskhod 0.222 · titan 0.173 · shuttle 0.116 · space 0.116 · thor 0.068 · agena 0.057 · discovery 0.034 · atlantis 0.029 · star 0.025 · columbia 0.024 |
| rocket | 4 | march 0.301 · long 0.301 · soyuz 0.185 · delta 0.089 · fregat 0.069 · medium 0.014 · heavy 0.013 · volga 0.006 · vanguard 0.004 · juno 0.002 |
How well it did
| Model | Coherence (c_v) | Log perplexity |
|---|---|---|
| mission names | 0.765 | -5.6 |
| rocket names | 0.746 | -3.49 |
Coherence around 0.75 is a reasonable result for short, name like documents. The mission name model scores slightly higher than the rocket name model, so its topics are the more interpretable of the two.
Limitations
What this analysis cannot tell you
- There is no predictive model here. LDA is unsupervised. Nothing on this site forecasts launch success, and no accuracy figure exists because no classifier was trained. Treat the topic model as description, not prediction.
- Cost is mostly missing. Price is absent for 3,365 of 4,630 records, 73 per cent, so no economic claim can be made from this data. That is why there is no cost analysis on this site.
- Operators are not the same as countries. The data records the flying organisation. Soviet and Russian activity is split across RVSN USSR, VKS RF and Roscosmos, so any national comparison has to combine them by hand, as the 44.5 per cent figure does.
- The record ends in 2022 and the final year is incomplete. Nothing here describes launch activity after that point.
- Success is coarse. Outcomes were flattened to a binary flag for rate comparisons, which counts a partial failure as a failure. 107 records are affected.
- Success rates are not comparable across eras. Each operator's rate is pooled over its whole history, so an organisation that flew mostly in the 1960s is being judged against one flying today, when the engineering is not the same. The confidence intervals capture sampling uncertainty only, not this.
- The permutation test is observational. It shows that experience and reliability are not associated among established operators. It cannot say why, and it does not rule out an effect within any single operator over its own lifetime, which this data cannot address.