Below I discuss nouns. I've looked at verbs just a bit in this same way, and have gotten nowhere. The short adjectives are the third territory, whose stress is a nightmare, and which likely is the least structured of all—or perhaps not.
What I have found, ultimately, is that the difficult and unpredictable stress-movements that occur are statistically demonstrable as only occurring in common words (the kind of finding everyone nods along with as if obvious only after it's been demonstrated, but whatever). This leads me to suspect that if I collected the data, the adjectives would be something like 90% predictable in the language, with just a few hundred odd short forms.
Anyway, to the point, here I have broken down noun stress using a dataset of the 10,000 most common nouns pulled from a database I produced from Wiktionary. Since the more rare words are extremely unlikely to have mobile stress or anything interesting to say, I would suggest that my numbers are essentially representative of the true value, within a very small margin of error. Supposedly a total high school vocabulary is around 20,000 words, so 10,000 nouns is uh... Like all of 'em.
A refresher on the model
Claim: Russian nouns have a very reliable stress structure, which splits singular and plural forms, that is, the singular forms all agree in stressed position, and the plural all agree. (No one ever mentioned that to me !!!!)
The stress position comes in two types:
(1) Early, that is, some characteristic syllable within the stem;
(2) Late (a reframing of the research's use of the term "desinential", for the following reasons), where the stress falls on the ending as much as possible. (This is different from saying "the last syllable", e.g. "-ами", of course, is always stress on the "a" when it is stressed (apart from the "ьми" stuff, I know, stay with me here).
I have devised the maximally compressed representation, seen in the following examples:
vocab [singular, plural]
кошка [early, early]
союз [early, early]
дом [early, late] (дом, дома, дому vs дома, домов, домах)
маяк [late, late] (note here why I use the term "late" rather than ending stress, since the zero-ending causes 'the next latest thing' to be stressed)
право [early, late] (note the genitive zero ending causes "as late as possible" stress, all according to the system)
дверь [early, late*]
сад [early, late]
широта [late, early] (широты versus широты, широтам, широт etc.)
This is the general rule for most nouns. Indeed, 91% of nouns have the same stress in all forms.
Of the ~10% of nouns where the singular stress and the plural stress differ (lots of the commonest nouns),. In these, ~77% of the time the plural stress jumps later in the word as in [early, late]. By contrast, 23% of the time are [late, early].
*Now... There are two caveats to the uniformity I've described of nouns' stress positions based on number (singular vs plural), which do not constitute exceptions to the system, so much as tweaks to what "the system" is. That is to say, natives do not find these to be "exceptions" to rules, unlike truly odd forms like "колец" (кольцо [late, early] gen.plur.exc.). They just know these tweaks deep down, and do not run afoul of them. If no one explained these tweaks, you would likely form a proper intuition for correct stress in all but the most common, wonky words. These wonky common words would be a forever source of unexplained confusion. Two forms can misbehave from the rest of their number. They are:
1) The singular accusative. (see statistical group 1 below)
2) The nominative plural. (see statistical group 2 below)
In my notation I put the asterisk beside the number which exhibits the exception. For example, above the asterisk by "late" for дверь denotes that it is nominative plural двери, apart from the rest of the plurals дверям, дверей etc.
As a result I have the following statistics from ~10,000 most common nouns (as determined by having processed dozens of books for frequencies).
(It should be noted that the stress of the uncommon locative case is sort of inherently on the ending, the paucal occurs in like 5 words with a notable stress of its own часа, and eh... the partitive so far as I can think only occurs unstressed....)
The Stats
The Acc Sing Feminines
1) There are 17 words, all feminine, where the accusative singular has earlier stress, differing from the singulars. (~.2% of 10,000 words) That's it. The accusative singular ever been a stress that didn't match the other singulars? It was one of these.
Can occur in these variations:
[late*, late]
[late*, early]
9 of these are simultaneously also part of group 2, explained below.
| sing | plur | |||||||||||
| nom | acc | gen | dat | ins | pre | nom | acc | gen | dat | ins | pre | |
Also note, that the accusative plural never really exhibits its own character. It either pulls from the nominative plural, or from the genitive plural, according to animacy rules. Because of this, it is simpler and more accurate to say that, in the plurals, the nominative plural is all that might misbehave, but the accusative plural can't be blamed for going along.
This word is considered to have a true irregular form only in the genitive plural. This is because the strange accusative singular is from being part of the magic 17: this is normal. However, the plural stress is "early" stress, yet for whatever reason the genitive plural has an unexpected late stress. Notice the uniformity for the singulars other than that. This word can be labeled: [late*, early] gen. plur. exception
This first group, the accusative singular exceptions, should be isolated in learners' minds to prevent their oddity from implying that any other word they ever learn might follow that pattern. These are the annoying Russian words for learners once they try free-flowing conversations.
The Early Nom Plur
2) There are ~154 (1.54% of ~10,000) words (some shown above), of masculine and feminine (and a few neuters), where the nominative plural is an anomalous early stress. Here is a group of them:
Comes in:
[early, late*]
[late, late*]
Note that groups 1 and 2 are the fundamental counter-expectational dynamics, and they both make for a stress that is earlier than what your intuition would have generated from the rest of the forms. Put differently, if the stress is already 'early' stress, it can't misbehave! This is why the [sing, plur] labels only ever have asterisk on a 'late'.
| sing | plur | |||||||||||
| nom | acc | gen | dat | ins | pre | nom | acc | gen | dat | ins | pre | |
They contain a lot of "-сть" ending nouns.
I have yet to find a real reason for these being like this, nor any kind of way to guess.
The Hidden Late Masculines
3) There are ~650 (6.4% of 10,000 nouns) masculine nouns with "hidden desinential stress", whereby the nominative singular has final syllable stress and a nominative zero-ending (the consonant, basically), which does not tell the observer whether or not the true singular stress is on that final nominative syllable ('early'), or on the ending ('late'). If the learner assumes the stress belongs where the nominative says, they will be corrected, in the oblique cases, that the stress falls actually on the ending itself. About 240 of these are fleeting vowel instances, which are reasonably detectable from nominative alone.
Comes in:
[late, late] (but you wouldn't know from nominative singular!)
That leaves 410 instances (as a slightly high estimate) of foot-gun style masculine nouns. (Truly not a huge number when you put it into a language learning context! They're actually rather few!)
| sing | plur | |||||||||||
| nom | acc | gen | dat | ins | pre | nom | acc | gen | dat | ins | pre | |
I have checked the data and found that among the hidden desinential masculine nouns, there are a couple with irregular "ья" plurals, and apart from that 100% of them exhibit stress always on the ending (which is to say, in the plurals too).
"If it's a masculine noun with hidden stress on the singular ending's syllable, all the stresses are late." For instance, "топор" with dative "топору" implies to a native that it is therefore "топорам." Since there are no nominative plural (and definitely no accusative singular) exceptions in these nouns, it also implies to them clearly it that it must be plural "топоры," no funny business.
A [early, late] masculine noun is normal. What I'm saying above is that a [late, ...] masculine noun where that late stress was hidden MUST be [late, late]
4) All the rest of Russian nouns, which are essentially normal, where the stress of the singular and the plural may or may not differ from each other, but each follows their internal order.
Comes in:
[early, early] (what you could just call "static" stress)
[early, late] (common)
[late, early] (less common)
[late, late] (for feminine or neuter nouns this is regular, without the difficulty of masculine zero ending)
Here, again, about 10% (~1,000 in dataset) of the time the two will differ, predominantly by a plural stress jumping later in the word, to the ending.
90% of the time in the dictionary, the stress simply doesn't move. That does not mean 90% of the time in speech of course, due to common words being more likely to misbehave.
The black middle column denotes differing stress between singular and plural
(You can ctrl+f in this)
When you see the charts on Wiktionary, sometimes some options are provided:
свеча is [late, late*] with a dated alternative realization as [late, early]
вода is [late, early] with a dated and obsolete realization as [late, late*]
I would recommend that students seeking to resolve their mis-stressing syndrome knock these groups out "in order" so to speak. Learn group (1), and rest assured that that oddity will never be seen in the rest of words they learn. Learn group (2), also very finite, and perhaps simultaneously learn group (3), knowing that the two are mutually exclusive. The following should minimize confusion for students:
+-----------+
| Group 1 | (17 very common basic crucial words)
+-----------+
|
v
+-----------+ +-----------+
| Group 2 | | Group 3 | (Two very finite groups of mostly useful basic words)
+-----------+ +-----------+
\ /
\ /
v v
+-----------+
| Group 4 | (The rest of the language, where singular and plural must be taught as нередко decoupled)
+-----------+