Exploring NFT Price Distribution Across Collections

NFT categories such as virtual land, PFPs, and game assets are a common framework to evaluate projects and collections. However, a less discussed and sometimes counterintuitive property of these assets is its price “tier” within the collection, and how assets of the same price tier behave across collections and NFT categories.

Gringotts DAO is on a mission to be a one-stop-shop for NFT holders to access liquidity. With new NFT financialization protocols rapidly emerging, we set out to evaluate the efficacy of different approaches in the context of what type of NFT the user is seeking liquidity for. Instead of focusing purely on the asset category, we looked into the property that all these assets share — their price.

More specifically, we sought to answer 3 questions:

• What does the price distribution of NFTs look like across the whole market?
• Are there price distribution patterns that emerge, and if so how common are they?
• From these distributions, how can we define price “tiers” that might render a given NFT more suited for certain liquidity methods over others?

One of the main findings was that items across collections and NFT categories may behave more similarly than those within the same collection. Therefore, price distribution analysis may give users and developers alike a more holistic view of where to best find liquidity, and the addressable market for financialization approaches, respectively.

Methodology

NFTBank is an algorithmic asset valuation product that uses machine learning to predict NFT prices based on past pricing of similar assets. We pulled data from NFT Bank over 3 months. First on 12/15/2021 (279 collections, ~2.4mn NFTs, ~3.7mn ETH market cap), then on 01/13/2022 (540 collections, ~14.2mn NFTs, ~8.9mn ETH market cap), and most recently on 02/27/2022 (538 collections, ~14.8mn NFTs, ~6.5mn ETH market cap)

This article dives into 4 observations we found:

1. Price distributions are generally very concentrated both across and within collections.
2. Price distributions come in 5 major “shapes” which do not seem to correlate to the NFT “category” (PFP, games, virtual land etc).
3. Price distribution shape generally remains constant. For 75% of collections, price distribution stayed the same across time points. For those that changed, it was toward a “related” shape.
4. For collections with exponential-decay and log-normal-like distributions (60% of collections), we can define and examine the behavior of floor-, mid- and top-tier assets.

Concentrated price distribution

Across collections, the market is concentrated in that the top 10 collections account for >60% of the market cap, with a (normalized) gini coefficient of ~0.9.

Within collections, most price distributions follow a pattern where most items are priced close to the floor. The few remaining items make up the majority of the price range and thus a significant contribution to the collections’ market cap.

Examples of normalized price distribution charts:

*Normalized price = (price — min_price)/(max_price — min_price)

In these charts, the x-axis is split into 100 equal parts, so the first chart (CryptoPunks) for example suggests that almost all Punks are priced within the first 2% of the full price range.

This is promising for NFT financialization products most suited for floor items, for example, liquidity pools like NFTX which can serve as “Floor AMMs” providing instant liquidity for NFT owners who can trade their floor assets against the pool.

Collections with a lot of floor items and a reliable price feed (those that are frequently traded between lots of different unique addresses) also make good candidates as collateral in P2Pool lending products. This is because floor assets can generally be treated “the same” and thus do not require a manual appraisal. Loan terms can be automated once plugged into a price feed and an automated means to evaluate risk.

In the above sample, note that some collections (VeeFriends and Decentraland for example) do not fit this “mode is the floor” pattern, however. In fact, the price distribution patterns fall into one of 5 distinct shapes, which brings us to our next observation.

Price distributions come in five major shapes

Across collections, the price distribution shapes we observed were:

• 1) Exponential decay. These are the collections with the majority of its items priced around the floor, with a long tail of higher-priced items. ~40% of the collections we sampled exhibited this profile. Examples include Cryptopunks, RTFKT Clone X + Murakami, and Mutant Ape Yacht Club

• 2) Lognormal-like distributions have a similar shape as exponential, but the mode is slightly higher than the floor price. ~20% of the collections we sampled exhibited this profile. Examples include Bored Ape Yacht Club, Sandbox LAND, and Decentraland.

• 3) Symmetric (or normal-like) distributions are where there is high asset concentration around the mean price, tapering off on either side. ~5% of the collections we sampled exhibited this profile. Examples include Anonymice, Blitmap, and Rollbots.

• 4) Multi-modal distributions exhibited several bumps and spikes across a wider range. ~20% of the collections we sampled exhibited this profile. Examples include VeeFriends, Autoglyphs, and FLUF world.

• 5) Point-distribution patterns have one of the above shapes, but with prices spread over <0.1 ETH. Because we defined this as being roughly the same price, we label these “point-distribution”. This shape is a common characteristic of smaller cap collections (except PVFD, none of the top 100 collections exhibit this shape) — so they function as a sort of filter. ~15% of the collections we sampled exhibited this profile. Examples are PVFD, Zodiac Capsules, or PEGZ.

It’s interesting to note that NFT categories (PFP, virtual land, game assets etc.) do not correlate with price distribution shapes. Virtual land NFTs in Cryptovoxels, Decentraland, and Somnium Space for example all have different distributions (exponential, log-normal- like (was symmetric in Jan/Dec data), and multi-modal, respectively).

It’s likely that price distribution is a function of inherent features of a collection itself, rather than the NFT category it belongs to. For land, this might be location, parcel size, foot traffic (revenue-generating potential), land that has already been built on and is thus sold at a premium, etc.

Next, we looked at whether these price distributions changed over time.

Price distribution (generally) stays constant

Due to limited data here (3 data points), only time will tell if the analysis here continues to hold into the future. Looking again at the normalized prices, we can see that price distributions from December (gray) and January (red) are often, but not always, congruent or at least similarly shaped as those from February (blue).

From the 537 collections that were included in both the January and February data pulls, 166 changed in price distribution shape (30%). For January to December, we also saw a similar portion change (25%). This may sound like a lot, but keep in mind that the above classification of collections to their distribution shapes is a bit fuzzy as we were not too stringent with cutoffs.

For example, one could make the distinction between exponential-decay and log-normal such that: “if mode > floor price => log-normal”. Looking at below ratios of mode to floor-price, we chose a more relaxed definition and allowed the mode to be even 10–20% above floor since we looked at the fitted distribution to classify their shapes.

Based on this, we consider exponential-decay and log-normal distributions to be “related”.

For cases where a change in price distribution was observed:

• ~42% changed to/from a point distribution. Point-distributions have one of the other four shapes, just on a very narrow price range
• ~26% changed from exponential decay or log-normal like to multimodal. This class is also defined more softly as our distributions typically have just one mode. We defined this shape though to separate distributions like VeeFriends with its several bumps (modes) from the other shapes.
• ~22% was exponential decay to/from log-normal like (if we’d go the stringent approach, this number would be way higher)
• ~10% remaining changes are all to/from a symmetric distribution, with log-normal-like distributions having the major share (6%). This is again due to a rather soft definition of the line between log-normal like and symmetric distributions (i.e. those two shapes are also “‘related”)

Defining price tiers

Based on the above observations, we looked to collections that had exponential decay and lognormal-like shape to define price tiers, as the floor price can serve as a reasonable anchor point here. Of course, since the “absolute” floor might be just one item with the lowest listed price, we wanted to find an appropriate multiplier under which to classify more items as a floor item.

Defining the floor: we looked at different lower quantiles and their ratio to the floor price.

For ~90% of these ~800 collections, the Median is below 1.4 * floor price. Picking a threshold here depends more on the use case we have in mind: if we go farther to the right to include a larger share of the collections’ items, we run the cost of extending its price range, making this set less homogeneous.

To get the threshold working for ~90% of collections, a threshold of

• 1.3 gives ~ the 25% quantile (so covers 25% of items)
• 1.4 gives ~50% quantile / Median
• 1.75 gives ~75% quantile

Less than 30% of a collection is likely too little, while a price range of [floor price, floor price*1.75] might be too wide. Therefore we opted for a multiplier of 1.4 for the floor cutoff. In other words, “floor tier” items are defined as those within the price range of [floor price, floor price * 1.4]. For 2/3 of the collections, this includes 75% of items.

Defining top items: we can follow a similar route using top quantiles:

A threshold of 2.5 covers 90% of the collection ~85% of those ~800 collections. It also includes 95% of the collection in 2/3 of those collections and even 99% for ~20% of those. In other words, a threshold of 2.5 would put the top 10% of assets in the “top-tier” bucket for 90% of collections.

Again, we can be more exclusive for this set and e.g. increase this threshold to 4.

Under these definitions for floor and top items, we can define mid-tier items as those with prices between [floor price * 1.4, floor price * 2.5]. We now turn to the characteristics of these price tiers.

Characteristics of defined price tiers

Floors. Items priced at floor price to floor price * 1.4.

Floors typically made up 50–75% of items within a collection, and 25–50% of its market cap. Their quantity and homogeneous behavior render them suited for liquidity pools, which effectively act as “Floor AMMs” where users can earn a yield on the trading activity of floor assets and enjoy the deepest liquidity compared to other price tiers.

Mid. Items priced floor price * 1.4 to floor price * 2.5.

Mids typically make up 20–40% of items, and 10–20% of a collection’s market cap. As things stand today, mid-tier items may be the least lucrative for trading, as they command less liquidity than floors, and have less exposure to reflexive upside compared to grails. Collections, where the mode is in mid-tier (those with symmetric price distributions), may be those where many users are more interested in the properties or utility of the asset itself rather than price. For example, virtual land floors might be too small or in unlucrative locations, while large, high-traffic land is too expensive or are not for sale. Thus, land buyers hunt for assets with good location and land dimensions as well as price.

If it turns out that mids contain some ‘transitory’ items, i.e. floors gaining or grails losing value, then this could be the tier for speculation and related hedging applications.

Top or “Grails”. Items priced > floor price * 2.5.

Grails typically make up ~5–10% of items, and 20–40% of a collection’s market cap. Grails have very noisy distribution and large price changes, behaving similarly to high-end items in “traditional” art or real estate. While their trading volume and velocity is low, grails have good potential for use as collateral or getting liquidity via fractionalization.

Regarding the share of items that fall in each of the three tiers we see a large share of floor items (blue). Here and there it’s rather small, but this ties back to the blurred definition of our shapes, e.g. Meebits (1st bar) doesn’t follow our tier logic fully as it has these additional bumps we showed further above:

Collection names are tiny, but a ‘(‘jan)’ or ‘(dec)’ at the end of the name tells it was from the January or the December data set respectively.

Things get a bit noisier when we look at market share for each price tier. Regarding market-cap share of those tiers, things get a bit noisier. While floors still seem to make up the majority of market share, it is common for collections to have grails that are 10–1000x above the floor, eating into the collection’s market cap.

Overall, ~25–50% of the market cap fall into the floor tier, 10–20% to the mid-tier, and 20–40% to the top tier:

Future Work

With this article, we took some preliminary steps to categorize NFTs by their price movement behavior and tier within their respective collections. As we alluded to above, tier cut-offs can be adjusted depending on the use case. For us, one of the purposes was to arrive at common behaviors and characteristics of NFTs across collections and asset categories that would inform holders of the best avenue to find liquidity. The analysis helped inform this evaluation matrix.

Now that we have a high-level overview of how assets behave across collections, we can zoom back into the notable observations we have made here and analyze it further. For example:

• What are the main properties of a given collection that might cause them to have the price distribution patterns that they do?
• What are the internal (e.g. project developments) or external (e.g. market sentiment) factors that would cause a given collection to change its price distribution shape over time?
• Could price distribution be a leading indicator or analysis metric for a given financialization protocol to onboard a given asset e.g. for collateral or to spin up an NFT AMM?

We hope to explore these questions in future articles. For now, we offer the quantified mental model for defining price tiers, and a preliminary framework against which to evaluate our assumptions for NFT liquidity methods in the months to come. If you have feedback on this article or further questions you want to tackle with us, please reach out to get in touch with the analytics guild at Gringotts.

*Note: pricing data are estimates by NFT Bank, not actual trades. So these figures should be taken with a grain of salt. For more information about NFT Bank’s approach to NFT valuation, refer to this article.