I've read a couple books on candlesticks, but they are all from at least 8 years ago. It seems like it has grown in popularity since then, and a lot of platforms offer the algorithms for people to use. Has this accessibility diminished the effect of these algorithms?
The difficulty with candlestick analysis is that the underlying data is inaccurate, particularly for thinly traded securities, with occasional blocks.
Large orders happen "off the tape," and are inserted into the record later. What happens is that if you put in an order for 10,000 shares, it may be made up of dozens of smaller orders. Those small trades never make it onto the tape. Instead, they are averaged together and reported after the final closing trade is completed at a point where it would not impact the market. If the high or low for the day happen to be in the block order then that information is lost as the average is the only thing reported.
A blip on the ticker you see moments before may have happened slowly over the last few hours. The only potential warning sign is the volume.
It also makes it look like trades are less frequent than they are. If you had a trade of 100, 100 and 400 shares over three minutes, evenly spaced, then it could look like 100,400 or 100,100, or 100, or 400 depending on which pieces ended up in the block order. It could also be 100, 100, 400 if none were used for the block order.
If you could filter out large orders from the ticker and restricted yourself to small trades, then you would have an actual time-series, though censored. Censoring is its own branch of statistics.
There are highly technical methods that could be used to estimate the impact of the block order but they go so far beyond the skills required for candlestick analysis that no one who could use them would likely use candlestick analysis.
For candlestick analysis to be useful, you would need an unusually high level of technical skill. A statistician could do it, some mathematicians could do it, and a handful of economists could do it. Because of the techniques required to perform it in real-time, you would probably have to code it in Rust, C or C++. That really shrinks the pool of people down.
While I personally suspect that some technical methods work in specific cases, I think most don't. I think of technical methods like surfing. You cannot surf on most beaches or rivers. Under certain virtuous circumstances, you can ride a surfboard. The rest of the time, you will fail.
Having read at least some of the academic literature on the topic, I don't think anyone really knows. The obvious suspects can be ruled out, but it isn't a literature that is broad or deep.
One last side note, if candlestick analysis worked it would have had to work "in the old days," when trades were thin, far and few between. Even two decades ago the NYSE had a post for stocks that only trade once every two weeks. When volumes were small, the record would have been more accurate. With things like program trading and other such developments, the candlesticks are not accurate.