Skip to content
guideOctober 31, 2026 · 12 min read

How to implement proof of reserves for an exchange, and what it cannot prove

Merkle sum trees, inclusion proofs, wallet signatures and per-asset ratios: how to implement proof of reserves, and what it cannot prove about solvency.

Kostiantyn Boiko, Blockchain Engineer

Proof of reserves has two halves. For liabilities, you commit every customer balance to a Merkle sum tree, publish only its root and per-asset totals, and give each customer a proof that their balance is inside. For reserves, you read the balances of your wallets at a recorded block and prove you control those wallets with signed messages. Then you compare the two per asset, sign the result and repeat on a schedule.

What it shows is narrow: on one date, the coins in your wallets covered the customer balances in your ledger. It does not show that you are solvent, that the coins were not borrowed for the day, or that anything is still true tomorrow.

We built a Proof of Reserves Kit that does this on the exchange’s own servers, and this guide uses its formats as a worked reference.

Why do exchanges publish proof of reserves?

FTX filed for Chapter 11 on 11 November 2022. The SEC later alleged that customer money had been diverted to Alameda Research while FTX advertised “best-in-class controls”. Customers could not check whether their balances were backed.

Within weeks, large exchanges published proofs. Binance released a Merkle-tree proof for bitcoin on 25 November 2022, reporting a 101% ratio against 575,742 BTC of customer balances; OKX’s first Merkle-tree report is dated 22 November 2022. Kraken now publishes a proof that it says includes total client liabilities, reviewed by an independent accountancy firm.

The backlash came as fast. In December 2022 Mazars, which had produced reports for Binance, KuCoin and Crypto.com, paused that work “due to concerns regarding the way these reports are understood by the public”. DLA Piper’s March 2023 summary of a PCAOB staff advisory reports that proof-of-reserves reports do not provide meaningful assurance to investors or the public and are not financial-statement audits.

What are the parts of a proof of reserves?

The same parts apply to proof of reserves for exchanges of any size, custodial wallets and OTC desks:

Part What you publish What it proves
Merkle sum tree Root hash, per-asset totals The totals add up the committed balances
Inclusion proof Privately, to each customer This balance is inside the total
Reserve readings Addresses, balances, block These addresses held these amounts at that block
Wallet-control evidence Address signatures or wallet-specific evidence Demonstrates signing capability or satisfies the stated control test, not legal ownership or exclusive control
Coverage Eligible reserves ÷ in-scope liabilities, per asset Shows the reported coverage ratio, which may be below 100%
Signed snapshot Snapshot data and a digital signature Detects changes to the signed data when checked against a trusted public key; does not prevent later changes to reserves
Attestation An accountant’s or auditor’s report Someone independent performed stated procedures

How does a Merkle sum tree prove liabilities?

The usual proof of reserves Merkle tree is a sum tree: a Merkle tree in which every node carries a hash and a balance. A leaf holds one customer’s balance. A parent’s balance is the sum of its children’s, and its hash covers both child hashes and both child balances, so the root commits to every balance and to the total.

The leaf contains a balance and a pseudonymous identifier rather than a plaintext name or account number. In our kit it is a SHA-256 hash of the snapshot ID, the user ID and a fresh random nonce per customer per snapshot that only that customer receives. Keeping a fresh, high-entropy nonce private makes identifier guessing and cross-snapshot linkage harder, but balances and other metadata can still reveal information.

A worked example with four customers

One asset, USDT, and four customers holding 500, 250, 1,000 and 50. H is SHA-256; distinct prefixes separate leaf and internal-node encodings. This is schematic notation: production code needs canonical field encodings, integer amounts in base units, asset and snapshot binding, and explicit bounds and overflow checks.

                 Root   sum 1,800
               /                  \
       N12   sum 750          N34   sum 1,050
       /          \            /            \
  L1   500     L2   250    L3   1,000     L4   50

L1   = H(0x00 ‖ uid1 ‖ 500)
L2   = H(0x00 ‖ uid2 ‖ 250)
L3   = H(0x00 ‖ uid3 ‖ 1,000)
L4   = H(0x00 ‖ uid4 ‖ 50)
N12  = H(0x01 ‖ L1 ‖ 500 ‖ L2 ‖ 250)
N34  = H(0x01 ‖ L3 ‖ 1,000 ‖ L4 ‖ 50)
Root = H(0x01 ‖ N12 ‖ 750 ‖ N34 ‖ 1,050)

The exchange publishes the root hash and the 1,800 USDT total in a signed snapshot. Customer 2 receives a proof with their balance, nonce and two siblings, and checks it:

Step Input Result
1 Snapshot ID, user ID, nonce uid2
2 uid2, balance 250 L2, sum 250
3 Sibling L1 (sum 500), left N12, sum 750
4 Sibling N34 (sum 1,050), right Root, sum 1,800
5 Root in the signed snapshot Hash and sum must match

If customer 2 is omitted, the exchange cannot supply a valid inclusion proof for that account. If it commits 200 instead of 250, verification against the customer’s independently checked balance of 250 fails; a verifier that simply accepts the supplied balance of 200 could still pass. Each customer checks only their own path, so the more customers check, the harder it is to drop anyone.

Two details matter. Both child sums go inside the parent hash, not only their total: Hu, Zhang and Guo showed in 2018 that the original design (the “Maxwell protocol”) let an exchange prove smaller liabilities than it had. And no balance may be negative: a fake account at −300 would bring the total down to 1,500, and only a customer whose proof happened to contain that −300 could notice.

How do you keep customer balances private?

The example shows the leak. Customer 2 learns that one other customer holds exactly 500 USDT and that the other half of the tree holds 1,050. In this one-leaf-per-customer example, the lowest sibling reveals another customer’s exact balance; padding, split leaves or other tree layouts can change what it reveals.

Mitigations, roughly in order of cost:

  • Hashed IDs with fresh per-snapshot nonces, to reduce identifier guessing and cross-snapshot linkage; distinctive balances can still identify customers.
  • Shuffled leaf positions every snapshot.
  • Splitting large balances across several leaves.
  • Zero-knowledge proofs, which hide other balances and prove properties of every leaf.

Get the hashing right too. Chalkias, Chatzigiannis and Ji (2022) found exploitable flaws in deployed proofs caused by SHA-1 or hashes truncated to 8 bytes, wrong Merkle inputs and missing user-ID uniqueness.

Merkle sum tree or zk proof?

Plain Merkle sum tree Zero-knowledge proof over the tree
Published Root hash, per-asset totals Root, totals and a proof about every leaf
Negative balances and fabricated accounts Inclusion proofs do not establish non-negativity or account authenticity across the tree Can enforce encoded balance constraints, but cannot establish that accounts are genuine or the source ledger is complete
What one proof reveals Sibling sums, including one other balance Little or nothing about others, depending on design
Customer inclusion check A logarithmic number of hashes, typically feasible in a browser Depends on the proof system and customer verifier; OKX’s repository separately estimates about 30 minutes with 8 GB of memory for its global verification workflow
Engineering risk Encoding, sum validation, privacy and ledger-reconciliation errors Additional circuit, prover and verifier complexity; some systems also require a trusted setup
Examples Our kit; Kraken uses a Merkle tree checked by an independent accountant Binance (Groth16 via gnark), OKX (its public proof-of-reserves-v2 repository uses Plonky2)

Contemporaneous reporting dates Binance’s zk-SNARK rollout to February 2023 and OKX’s zk-STARK rollout to April 2023. Binance describes a non-negative total-net-balance constraint, while OKX documents balance, non-negativity and inclusion constraints. These apply to the committed data under each system’s definitions, not necessarily to every individual asset balance, and do not establish ledger completeness. For a small or mid-size exchange, a sum tree with nonces, shuffling and an auditor who checks for negative leaves is a reasonable start; zk proofs are the upgrade once that is routine.

How do you prove the reserves?

Read balances at a recorded block. Inventory every address holding assets included in reserves, including unswept deposit addresses; use a documented completeness check rather than making their inclusion optional. Read native coins and tokens, and record the block height and hash of each reading. Pinning differs per chain: EVM chains let you query every balance at one block number; on Bitcoin, check the tip before and after reading; TRON’s HTTP API serves only current state, so record the head block when the read starts. A wallet you could not read is listed, never counted.

Prove control with signed messages. The auditor supplies a fresh nonce, and the authorised wallet-signing workflow signs a challenge bound to the company, network, address, snapshot and an expiry time. Keep private keys out of the proof-of-reserves service. Our kit’s message:

BitPolaris proof of reserves
Company: <company>
Challenge: <challenge id, e.g. 2026q4>
Address: <address>
Nonce: <auditor nonce>

Single-key wallets sign with EIP-191 personal_sign on EVM chains, signMessageV2 on TRON and signmessage on Bitcoin. That classic Bitcoin format was built for legacy 1… addresses; BIP-322 generalises it, and EVM smart-contract wallets verify through ERC-1271. Until the verifier supports the relevant wallet types, classify them separately and use an auditor-agreed control procedure. A witnessed transfer can support that procedure but does not by itself prove ownership, exclusive control or freedom from encumbrances. OKX publishes “I am an OKX address” signatures, so anyone can check them.

Count only what is yours to pay out. Assets at a custodian or another exchange go in from statements or a read-only API, labelled as such. Exclude assets unavailable to meet customer withdrawals from eligible reserves. Disclose borrowed assets and the corresponding repayment obligations separately; a wallet balance alone cannot establish that assets are unencumbered or that related debts have been included. Your custody policy has to keep them apart, and your wallet and approval model decides how easily an auditor can see that.

When should you take the snapshot?

A proof of reserves describes one moment. A point-in-time report may miss temporary borrowing. DLA Piper’s summary of the PCAOB staff advisory identifies borrowed assets as one of the matters such procedures may not resolve.

Timing raises questions even when nothing is wrong. In October 2022, Crypto.com sent 320,000 ETH to a Gate.io address; its CEO said it was meant for a new cold wallet, and the funds came back. Gate.io published its proof of reserves a week later and said its snapshot predated the transfer. A published snapshot time and block height settle that kind of question.

What reduces the point-in-time problem:

  • Snapshot on a fixed schedule, daily if you can, and publish the history, so reserves that appear only on report day stand out.
  • Take liabilities and reserves together: ledger export in one read-only transaction, chain reads at the next block, both times recorded.
  • Let the auditor select additional dates at short notice. Agree the procedure and notice period in advance; apply any jurisdiction-specific requirements separately.
  • Record what is in flight: pending withdrawals and unswept deposits.

How should you report the reserve ratio?

Per asset, never as one dollar figure, which lets a surplus in one coin hide a shortfall in another. Kraken and OKX both publish per-asset ratios.

Report eligible reserves per asset, with separate breakdowns for cryptographically verified wallet control, auditor-reviewed manual evidence and third-party statements. Do not treat every readable address balance as a reserve asset or a control signature as proof of ownership. The following figures are illustrative, not exchange data:

Asset Customer liabilities Reserves (readable) Ratio Reserves (ownership verified) Ratio
BTC 412.5 431.2 104.5% 418.0 101.3%
USDT 18,400,000 18,950,000 103.0% 17,100,000 92.9%
TRY (bank) 6,200,000 6,350,000 102.4% statement only —

The USDT line needs explaining: here, a multisig cold wallet awaits the auditor’s confirmation of manual evidence. List the outstanding control check as an exception, alongside unreadable wallets, any negative-balance adjustments and assets below 100% coverage. Explain each adjustment and exclude unsubstantiated assets from the eligible-reserve ratio.

Self-published or independently attested?

Self-published proofs let anyone check the arithmetic and their own inclusion, but not that every account is in the tree or that no debts sit elsewhere.

An attestation adds an independent party who checks the ledger export against the tree, tests ownership and reads the reserves. It is still not a balance-sheet audit: the PCAOB’s 2023 advisory notes that such procedures likely do not address the entity’s liabilities.

Some regulators now set the procedure. Law-firm and news summaries of SPK’s 8 May 2025 principle decision describe quarterly reserve audits for platforms within its scope, performed by independent information-systems audit firms, with two randomly selected comparison dates and wallet-control checks. Confirm the operative SPK text, scope and any later amendments before using those summaries as an audit specification. MiCA Article 75 requires a register of client positions, statements at least every three months and on request, and asset segregation for providers within its scope. That article does not itself require a public Merkle-tree proof of reserves; such a proof does not replace its custody obligations. That is our reading, not legal advice; what regulators ask for covers the rest of the technical file.

What can proof of reserves not prove?

  • Solvency. The tree holds customer balances only. Bank loans, payables, taxes and claims against the company are outside it.
  • That every customer is in the tree. An omitted account is caught only if its owner checks.
  • That reserves are unencumbered. Pledged or rehypothecated coins look the same on chain as free ones. Address balances and signatures do not establish whether contractual claims restrict the assets’ availability.
  • Exclusive control. A signature proves someone can sign, not that nobody else can. Other parties may hold copies of the keys or share signing authority.
  • Fiat. Bank balances rest on statements and on whoever checked them, not on cryptography.
  • Tomorrow. Keys can be lost or funds stolen the day after the snapshot.

How to implement proof of reserves: a practical plan

  1. Inventory. Every asset with its decimals, every network, every wallet by tier, every custodian, venue and bank account. A wallet missing from the list does not exist for the proof.
  2. Liabilities export. One row per customer and asset (user_id, asset, balance), locked and in-order funds included, as decimal strings, never floats. Use a transactionally consistent, reproducible export reconciled to the ledger’s control totals; a CSV is only a transport format, not evidence of completeness. An asset in the ledger but not in your configuration should stop the run, not vanish.
  3. Build the tree. SHA-256 with separate prefixes for leaves, nodes and padding, both child sums in each node hash, per-customer nonces, shuffled leaves. Negative balances go to zero in the tree and onto the exceptions list.
  4. Read reserves at a recorded block, per chain.
  5. Collect wallet-control evidence bound to the snapshot and a fresh auditor challenge. If evidence is reused, disclose its date and refresh it after changes to addresses, keys or signing policy.
  6. Sign the snapshot: root, totals, readings, signatures, bank balances, coverage and exceptions in one canonical file, signed with a dedicated key.
  7. Publish at a fixed URL such as /por/: the latest snapshot, its signature, the public key and the history. Serve each customer’s proof inside their logged-in account, because it contains their nonce.
  8. Archive daily snapshots for at least a year, and be able to take a fresh one on the day the auditor names.

How customers verify. They download their proof from the account area and open your verification page. It checks the snapshot signature against your published key, rebuilds the leaf from their balance and nonce, walks the path and compares the root with the signed snapshot. Publish the verifier’s code so technical customers can run it offline.

What the auditor gets. The signed snapshot, wallets with block heights and hashes, ownership signatures and manual evidence, bank statements and the exceptions list, and controlled access to verify sampled customer proofs, plus retained responses or suitable historical-node access to reproduce reserve readings. A current-state API cannot necessarily reproduce an earlier reading.

Proof of reserves FAQ

Is proof of reserves the same as an audit?

No. A financial audit covers the whole balance sheet over a period and ends in an opinion; a proof of reserves is a set of checks at one moment.

Does proof of reserves prove an exchange is solvent?

No. It reports reserve coverage for specified customer balances at a stated cutoff, and coverage may be below 100%. Solvency also depends on other assets, liabilities and obligations.

What is proof of liabilities?

The half of proof of reserves that commits to what the exchange owes customers, usually as a Merkle sum tree, so each customer can check their balance is counted in the published total.

What is a Merkle sum tree?

A Merkle tree whose nodes carry a balance as well as a hash. Each parent’s balance is the sum of its children’s, so the root commits to every balance and the total.

Do we need zk proofs?

Not to start. A sum tree with hashed IDs, nonces and shuffling, checked by an auditor, covers the basics; zk proofs add privacy and rule out negative accounts, with more complex code.

How often should an exchange publish proof of reserves?

Daily snapshots are cheap once automated. Publish on a fixed cadence, and let your auditor pick extra dates at short notice.

Does this apply to custodial wallets and OTC desks?

Yes. Anyone holding client balances has liabilities to commit and wallets to prove. With few clients, take extra care with leaf privacy: a sibling balance is easier to guess.

Proof of reserves with BitPolaris

The BitPolaris Proof of Reserves Kit runs on your servers, read-only against your ledger and the blockchains. It holds no wallet private keys, signs no blockchain transactions and never writes to your ledger. Snapshot signing uses a separate Ed25519 key controlled by your organisation.

  • Every day at 00:05 UTC it commits all customer balances to a Merkle sum tree and publishes only the root and per-asset totals; customers check their proof in the browser, on a page in your brand.
  • It reads wallets on EVM chains (Ethereum, BSC, Polygon, Avalanche C), TRON, Bitcoin (including xpub wallets), Solana and the XRP Ledger at a recorded block, plus Fireblocks and Binance accounts through read-only API keys.
  • It verifies signed ownership messages from single-key EVM, TRON and Bitcoin wallets; multisig, contract and custodian wallets stay “not verified” until the auditor confirms manual evidence.
  • It writes a snapshot signed with your Ed25519 key and an auditor pack: an HTML report in English and Turkish with CSV annexes, ordered like Türkiye’s reserve-audit list. Snapshots are kept for 400 days.

Not built yet: leaf splitting, zero-knowledge proofs, BIP-322 and ERC-1271 verification, and on-chain anchoring of the root. The kit is in pilot: it works end to end, and we install it ourselves with each of the first platforms. BitPolaris supplies software, not custody, payment processing or audit services. The auditor pack supports your auditor’s work; it is not an assurance opinion, regulatory approval or substitute for your licences and compliance obligations.

Our Custody software already reconciles the ledger against the chain, with a daily Merkle export for proof of reserves. The kit is €1,700 one-time plus €85 a month as an add-on for BitPolaris customers, or a €9,800 licence on your own servers plus 20% a year for updates. See the Proof of Reserves Kit, or tell us your ledger, chains and audit date.

Sources

Product names are trademarks of their owners. Descriptions of third-party products come from their public documentation and imply no affiliation.

Self-hosted custody software

Vaults, M-of-N approvals with hardware keys, whitelists and a reserve monitor that checks the ledger against the chain, on your servers. From €15,400 one-time.