Answer time time in statistics much longer than it actually is

Hello everyone.

The title basically says it all. In the last few days my answer time for some cards has increased four to five times. So not really. It still takes me about ten seconds per card. But according to statistics it is 40 to 50 seconds. Which is definitely not true. The phenomenon only takes place in a specific deck. And not always either. With one and the same card there are learning steps with a realistic value. And then some learning steps get completely out of hand.

What could be the technical problem? Is this known? Is there any solution to this?

Your help is very much appreciated.

Can you post –

  • Those stats that seem wrong
  • The Card Info (Statistics - Anki Manual) for cards that have stats problems and the Deck Options for that deck
  • The Card Info for cards that have learning-steps problems and the Deck Options for that deck
1 Like

Well, the best way to see this is in the Card Info under Time. I’m attaching three examples. One with 16 minutes of learning time is particularly strange…


And here are the associated deck options…

1 Like

The wrong statistics are then just a consequence. When I said it took an average of more than 30 seconds per card on the first day, I of course noticed this.

1 Like

The first thing I notice is that your max answer is set to 999 seconds = 16.65 minutes. If you took a break during your study session, that could absolutely be the time that was recorded. Unfortunately, AnkiMobile doesn’t display the timestamp for the review, but it would be interesting to see in AnkiDesktop – what time were the reviews just before and just after that.

And your learning steps / intervals appear to have some problems, but I assume I can’t see all of them based on the warnings you’re getting. Could explain more about what learning-step problems you see?

1 Like

I actually have a suspicion about the 16.65 minutes. It has already happened to me that I ignored the charge level of my battery for too long and then the cell phone shut down while I was studying. If I let it load before I continued studying, the minutes naturally add up. But that’s just a side curiosity anyway.

My Learning Steps:
1m 640m 16h 1d 36h 54h 81h 7290m 10935m 16403m 24604m 36906m 40596m 44656m 49121m 54034m 59437m 65381m 71919m 79110m 87022m 95724m 105296m 115826m 127408m 140149m 154164m 169580m 186538m 205192m 225711m 248283m 273111m 300422m 330464m 363511m 399862m 439848m 483833m 532216m 585437m 643981m 708379m 779217m 857139m 942853m 1037138m 1140852m 1254937m 1380431m 1518474m 1670321m 1837353m 2021089m 2223197m 2445517m 2690069m 2959076m 3254983m 3580482m 3938530m 4332383m 4765621m 5242183m 5766401m 6343042m 6977346m 127918h 8442588m 9286847m 10215532m 11237085m 12360794m 13596873m 249276h 16452216m 18097438m 19907182m 364965h 24087690m 26496460m

I don’t think there is a problem here. The cards don’t graduate at all. Above all, I’ve never had a problem with it. And I’ve been doing it this way for a very long time.

And to immediately answer the question of why. I want to be able to see my success rate at each step. This is only possible with an add-on. And there only in the Learning Steps.

My real problem is that learning times of 40, 50, 60 seconds and more suddenly appear. But that doesn’t correspond to reality. Of course anyone can say that. but it is like that.

I’m wondering, could a weak battery mess up the timing?

At the high end, your learning steps are equal to 41.6, 45.8, 50.4 years. These are purely theoretical, unless you plan to live for many hundreds of years.

You list the learning steps and say “I don’t think there is a problem here. The cards don’t graduate at all.” But in your initial message you said, “And then some learning steps get completely out of hand.” So it’s not clear if you feel there is a problem or question here. It’s certainly very unusual.

If you question is limited to answer time, the high time limit certainly seems likely to be the cause. One of your screenshots shows times for one card of 16.65m = 999s = your “Maximum answer seconds value”, 4s, 5s, 23s, which averages out to 257.75 seconds.

You use the terms “answer time”, “learning time”, and “learning steps”.

Just to be clear, are you using “answer time” and “learning time” interchangeably? In other words, this is the value show in Stats under the “Today” section in the line that says “Studied NNN cards in NN.N minutes today (N.Ns/card)” ? Whereas “learning steps” is what is specified in the deck options…

I do not understand the reasoning behind setting such a high maximum value for timer if it almost always takes you around 10 secs for a card. I set mine at 10-20 secs and 60 secs for my Kanji writing cards. Do you actually sometimes end up taking 16 mins to answer a card?

1 Like

Seems likely the issue. It’s hard to tell that it’s “just a side curiosity” to you, since it’s what you wrote your post about. :person_shrugging:t4:

I honestly don’t know what to say to that. You didn’t ask for advice about that, so I’m not going to give it.

What I will say, is that your settings are not in harmony with each other. I’m looking at you, minimum-interval and graduating-interval …
If you want to setup an unusual system, you avoid the most surprises if you at least make it consistent.

We can’t help you if you don’t show us those cards.

On the learning steps? No.

You’re probably right. The 16.65 minutes indicates the maximum learning time. And so that the card was not closed. I did a little more research and found the 16.65 minutes on other cards too. This seems to be where the problem lies.

I didn’t want to believe it at first because it seemed like a one-off event to me. I also kept seeing these learning times of 30, 40, 50 and more seconds. Even this length of time seemed unrealistic to me. Since these comparatively slightly increased learning times occurred much more often, I have located the problem here. In fact, learning times of just 5 seconds are even more common. These probably compensate for the 30, 40, 50 and more seconds. I made a mistake and your first impulse was probably right. The 16.65 minutes are the problem. Thank you for pointing out to me that 999 seconds equals 16.65 minutes.

I wanted to object. But I just see that I have selected cards with designated learning times of 28, 27, 33 seconds. Of course that’s not 40, 50 or 60 seconds. In this respect, these cards do not support my information. Since the other learning times on the cards posted are in the single-digit second range, I think the cards were still able to make my point clear. Doesn’t matter. I think it makes no sense to post cards with 40 and 50 seconds of learning time again. I realized that my problem lies with the 16, 65 minutes.

That would interest me too. As far as I can see, AnkiDesktop offers me the opportunity to see which card was learned today and at what time. The cards I learned today are shown in the overview but not in the order in which they were learned. Then the comparison with the card that came after the card with the 16.65 minutes will probably not be possible. Or have I overlooked something?

The problem I have now is that the iPhone has already switched off due to a weak battery. But not all the time. It doesn’t explain that I’ve been constantly having these cards with 16.65 minutes of study time for a few days now. You exposed the problem. Anki doesn’t seem to close the card properly and time continues running in the background. But this has to happen in other cases than when the iPhone turns off because of a weak battery. And strictly speaking, it shouldn’t even happen in a crash. I’m trying to test this out a bit now. First result: The further assumption that a crash is enough if the app is open in the background. Mind you, just the app. No card. That doesn’t seem to be a problem.

The harminia problem. That’s right. It would probably have been perfect to set the graduating interval, easy interval and minimum lapse interval so high (for many years) that Anki would probably not display a warning message. But it seems to me that there is a risk involved in doing this afterwards. Similar to how Anki can’t handle it if I change the learning steps later. The settings have been with me for many months and there has never been a problem. Nor should it. The cards don’t graduate. Then the cards cannot lapse and therefore do not need a minimum lapse interval. But of course I understand your basic criticism. I’ll take this into account when I create new decks.

That’s correct. The idea is to be able to keep the contents of a card for a lifetime. I didn’t take into account that the distances between all learning steps have to be added up when adjusting for lifespan. Then of course it takes fewer learning steps to reach 50 years. I can’t say exactly anymore. I guess I didn’t think this through carefully enough at the time.

Nonetheless. It doesn’t seem sensible to me to change that afterwards. Anki regularly has problems with changing the learning steps afterwards. And it shouldn’t bother if there are too many…

By Learning Steps out of hand I meant Answer Time of Learning Steps out of hand. So the Answer time is the problem. Sorry if that wasn’t clear.

I refer you to my previous post. I have now seen that.

No. Of course not. Maybe it can last over a minute if I’m distracted. Or the card is very complicated. I suspect I can easily change the Maximum Answer Seconds later. Then I could minimize the problem. However, I would still rather solve it. So understand why Anki occasionally doesn’t close the card and Answer Time continues.

At that time I set this to 999 seconds (maximum value) so as not to theoretically come into conflict with the Maximum Answer Seconds at all.

Now that you clarified for sprvlcn that you’re not asking about learning steps/intervals, I understand this was my mistake. Yes, 16m65s spent on one card would have an outsized effect on your average times.

I’m not sure this would be useful anymore, since we’re not trying to rule-out an issue in the time-per-card figures (which would have been made clearer if we saw overlaps between cards). But yes, you can see this in Desktop. In the Browse window, run a search for rated:1 (for today, rated:2 for yesterday, etc. – Searching - Anki Manual), then show and sort-by the “Card Modified” column – Browsing - Anki Manual . If you pop open the Card Info next to that, you can quickly scroll through your cards in the order they were (most recently) reviewed.

I’m 100% confident that no one has ever tested the edge case of “what happens if you have 80+ learning steps, with all cards in Learn, and change intervals?” – so I guess I can’t endorse that being low-risk. :sweat_smile:

I think it’s worth noting though – in the system you’ve constructed, you’ve essentially eliminated any chance of Anki behaving like the spaced-repetition system it is intended to be. You’ve chosen specific intervals (that increase at a plodding pace), and you force every card to maintain those. It sounds like that’s what you want, but it seems like a waste of so much time and so much of Anki’s potential.

To try and force every single flashcard go through similar learning steps is just so much of a waste of time. To quote Sherlock, “Different memories fade at different rates.” So it doesn’t make sense for everything to go through the same steps. Another drawback is if you fail something it goes back to the first step which would just create unnecessary extra work.

I’m really curious why do you need to know your retention at each step. Is it purely out of curiousity?


Rotate your phone to landscape, and it will appear.

The main reason for this setting is to keep your statistics sane when you get distracted and leave the card shown. The higher you set this, the less likely the averages will accurate reflect the time you take.


First of all, I think Anki is by far the best learning system currently available. But there is still a lot of room for improvement. Anki’s success is primarily due to the weakness of its competitors. There’s still plenty of room for a new champion.

Compatibility with Excel is great. So now I can work through a textbook. Properly structure the information in the book, which is only superficially structured as an alibi but is actually chaotic, in an Excel script. In the shape:


Then - depending on the level of in-depth knowledge - I only click on the information that I want to learn. VBA puts together the necessary flashcards for me. Taking into account the higher levels of knowledge that cannot be missed. Import. Complete. This means I can gradually deepen each learning material at different points. And never lose the connection in my available knowledge. No island knowledge is created. Terrific. Thanks Anki.

I don’t know of any other program that allows endless repetitions. Until the end of your life, so to speak. Technically a small thing. However, after a certain number of repetitions, all programs usually decide that the learning material has now been transferred to long-term memory. And therefore available for all time. The student can stop repeating the learning material. The approach is ridiculous. The brain also throws information out of long-term memory when it feels the information is no longer needed. This may not matter if I use an information regularly after learning it. Then the repetition happens intrinsically. However, people definitely want to keep information available that they do not need regularly. This makes such a finite system impossible. For me, this is the second major benefit of Anki.

The Spaced Repetition System that Anki is so proud of. Is a good thing. Answer pages can vary in difficulty. Which just means they carry more individual information. But that cannot be adjusted so easily. Twice as much information means twice as likely to lose the total information. Reducing the repetition interval makes sense. Well done, Anki.

But you forgot the much more important point 3. That’s why it has been submitted later and recommended for implementation

We like to locate our abilities and therefore our success in our intelligence. Really? Neurology and psychology have around 130 definitions of intelligence - if I remember the number correctly. They then all develop their own test, each of which leads to different results. And to top it all off. If the same person runs the same test concept several times, there will be different results. If there is such a thing as intelligence. Then one thing is certain. We have no idea what it is. Scientists know and accept this. Normal people who want to apply even partial knowledge as quickly as possible don’t do that. Instead, they carry intelligence in front of them like a monstrance as an explanation for their own success - if it exists, which they themselves don’t know where it comes from. I am convinced that the reason for conventional success lies in the amount of information that my brain does not have to look up but can reproduce by heart. I have a degree in economics and a degree in law. Every supposedly complicated train of thought that I encountered during my studies was only complicated because the information on which the train of thought was based was not available. Or the information in the train of thought was presented in an unstructured manner. Compared to simple school books, which are written exclusively from a didactic point of view and are constantly being further developed, university textbooks are a horror.

Every brain learns by heart differently. The difference lies in the interval between repetitions that a brain needs in order not to forget. Secure. Calculated across all people and all repetition levels, there is an average. Life is normally distributed. Maybe that’s the 250% that Anki claims. This is of no use to all brains above and below the average value. This is of no use to all brains with ease percentages that fluctuate depending on the repetition level. Both need individual values so as not to waste a lot of time.

You can set the average value in the options under Starting Ease? Could you. If only they knew where their personal value lay. They could determine this in a series of tests. However, Anki does not really provide the information that would be necessary for this. This is only possible with difficult work arounds. For this I refer you to the following article: Anki's Starting Ease Factor Setting.

But that’s not enough. There is simply no single ease percentage for increasing the interval that works permanently. If I can see the percentage of lost cards at each interval, it is possible for me to adjust the ease percentage based on how many repetitions I have reached.

An example: I achieve a 98 percent success rate in the first month with a starting ease of 150 percent. In order to maintain this, a reduction to 110 percent starting ease is required after a month. Otherwise I’ll suddenly lose 10 or more percent of the cards. And now? The Anki algorithm does not provide for this level of individualization. If a certain brain always functions at the same percentage, it doesn’t matter. It will point out its intelligence and make it feel superior. His success is only due to a ridiculous coincidence. Just as a photographic memory laughs at any other brain that uses repetition at all.

So how do I get the coveted percentages? By using an add-on that shows me the loss percentage for each individual learning step. Unfortunately, this only works at fixed intervals. And that’s only available in learning mode.

So what I do is this. I trade the advantages of spaced repetition for the advantages of “individual ease percentage per repetition level”. Because I think this is the much more fundamental adaptation to my brain. Spaced repetition is adaptation to the learning content and is good. But only if it is based on the “individual ease percentage per repetition level”. If I’m forced to make a decision, I’d rather forego spaced repetition altogether. My system is a work around with which I try to compensate for the adjustment that Anki skipped using on-board resources.

Again: For a brain that happens to function at the same ease percentage, this doesn’t matter. Just as repetition doesn’t matter for a photographic memory. But I bet the vast majority of brains out there aren’t that linear. Maybe that’s the only difference between the law student who uses Anki and the boy who has to deliver pizza to him in the evening because he doesn’t even have a decent school diploma. Of course the law student is now offended. He is firmly convinced that the difference between the two lies in his hard work, his discipline or his superior intelligence.

Nope. May it does not.

I understand the cleanup function. I will correct the value. But first I’m trying to figure out why this has been happening so frequently for the last few days. I don’t allow myself to be distracted that often. :wink:

@nepomoc Doesn’t FSRS does the same thing? You optimize every month or so in FSRS.

Anyway, I will just mention I don’t think if you simply stop using something you forget it. This only happens if you are not building up on a foundation. If the knowledge does not have any branch to hang on. But I really don’t know about things that you never use in real life. Well this is not a place to discuss such things so I’ll not. Though I’d say I would like to know how successful your approach has been(how the heck do we measure it though?)

I personally used to keep the ease very low. Almost near 130, the natural floor. If I notice problems with Retention any month I will just so slightly adjust the interval modifier. In your case, you are looking at individual learning steps if I’m correct. Wouldn’t FSRS do that for you though? FSRS reminds me, wasn’t there something called RMSE(bins)? Maybe It can potentially be used to measure how successfully your approach has been. Don’t know though.

Edit: I haven’t read it completely but I agree with some of the ideas in the blog. I assume it’s your blog. What you said in another blog post about lapses, I think FSRS solves the problem because how we can actually set it up such that you are shown the card once again when retrivability falls to 85 per cent.

1 Like

I just looked at FSRS again. I am not sure. But as far as I understand it, the calculation of the interval for a card is based on the data of the individual card itself. That would be exactly the same problem as described in the linked article. But perhaps it is actually the case that the experiences of all cards are taken into account when calculating the interval for a card. I’d have to ask the developer about that. Anyway, thank you for reminding me of FSRS again.

I don’t even want to think about it. At least for now I know it will take me a fraction of the time per card than I did before. And that with a success rate of 98 percent. I’m happy with that for now.

That would be too much of an honor. I’m just a reader too. :wink: