Opinion for the court filed by Circuit Judge MAYER.
Opinion concurring in-part and dissenting in-part filed by Circuit Judge DYK.
MAYER, Circuit Judge.Hynix Semiconductor, Inc. and Hynix Semiconductor America, Inc. (collectively “Hynix”)1 appeal the judgments of the Court of International Trade, which upheld several aspects of the final antidumping determination in the seventh administra*1366tive review of dynamic random access memory semiconductors (“DRAMs”) from the Republic of Korea. Hynix Semiconductor, Inc. v. United States, 248 F.Supp.2d 1297 (Ct. Int’l Trade 2003) (“Hynix I”); Hynix Semiconductor, Inc. v. United States, 295 F.Supp.2d 1365 (Ct. Int’l Trade 2003) (“Hynix II”); Hynix Semiconductor, Inc. v. United States, 318 F.Supp.2d 1314 (Ct. Int’l Trade 2004) (“Hynix III”). The Department of Commerce (“Commerce”), and intervenor Micron Technology, Inc. (“Micron”), cross-appeal. We affirm-in-part, reverse-in-part and remand.
Background
This case involves the seventh and final administrative review of the antidumping duty order on DRAMs from Korea, which encompassed the Period of Review (“POR”) between May 1,1999, and December 31, 1999. Dynamic Random Access Memory Semiconductors of One Megabit or Above From the Republic of Korea; Final Results of Antidumping Duty Administrative Review, 66 Fed. Reg. 52,097 (Oct. 12, 2001) (“Final Results”).2 Dissatisfied with Commerce’s methodology for determining the duty, which was set at 2.92%, Hynix appealed the Final Results to the Court of International Trade.
Hynix first argued that Commerce should have used a hybrid sales-plus-entries methodology for calculating the duty instead of using only sales completed during the POR. Second, Hynix claimed that Commerce should have allowed the amortization of its research and development (“R&D”) costs regardless of the fact that Hynix had previously expensed such costs. Third, Hynix disagreed with Commerce’s decision to use all R&D expenses, as opposed to product-specific R&D expenses, to calculate the duty. Fourth, Hynix argued that it should have been allowed to indefinitely defer certain R&D expenses as allowed by Korean Generally Accepted Accounting Principles (“GAAP”). Fifth, Hy-nix disputed Commerce’s refusal to offset foreign currency translation losses with the revaluation of fixed assets.
In Hynix I, the court affirmed Commerce’s decision to use sales during the POR to calculate the duty. It found that Commerce had supported its decision by showing that the sales-based methodology conformed to both standard practice and the practice in the preceding reviews. The court further found that Commerce had explained its decision to abandon the hybrid sales-plus-entries methodology that was used during the preliminary review. See Dynamic Random Access Memory Semiconductors of One Megabit or Above From the Republic of Korea; Preliminary Results of Antidumping Duty Administrative Review and Notice of Intent Not to Revoke Order, 66 Fed. Reg. 30,688 (June 7, 2001) (“Preliminary Results”). The court also affirmed Commerce’s refusal to offset Hynix’s foreign currency translation losses with the revaluation of its fixed assets. It agreed with Commerce that the revaluation of fixed assets did not represent income and could not, therefore, be used to offset the real losses incurred as the result of holding dollar-denominated debt.
Conversely, the court remanded the three issues relating to Hynix’s R&D expenses. First, it held that Commerce had failed to support its decision disallowing the amortization of R&D costs. On re*1367mand, Commerce was instructed to explain why Hynix’s amortization of R&D costs, which conformed to Korean GAAP and had been verified by Commerce, did not accurately reflect Hynix’s actual expenses. Second, the court remanded Commerce’s refusal to accept product-specific R&D expenses, which had also been verified by Commerce. Third, the court remanded Commerce’s refusal, which was based on the theory of conservatism,3 to allow the indefinite deferral of certain R&D costs.
Commerce’s remand decision closely followed its decision in the Final Results. See Final Results of Redetermination Pursuant to Court Remand: Hynix Semiconductor, Inc., Hynix Semiconductor America, Inc. v. the United States and Micron Technology, Inc., (Court No. 01-00988) (June 6, 2003) (“Remand /”). Commerce continued to hold that Hynix should not be allowed to amortize R&D costs because it had expensed such costs in previous reviews. According to Commerce, the change in accounting methods, while allowed by Korean GAAP, distorted the costs of production during the POR. Commerce also persisted in refusing Hynix’s product-specific R&D costs, holding firm to the theory of cross-fertilization.4 Finally, Commerce maintained that Hynix had not offered sufficient proof of future income to justify the indefinite deferral of certain R&D expenses.
In Hynix II, the court again affirmed-in-part and remanded-in-part. The court affirmed Commerce’s refusal to allow the indefinite deferral of certain R&D expenses. Specifically, it agreed with Commerce that Hynix had failed to provide documentation supporting its claim that the deferred R&D expenses would produce future revenues. The court, however, again remanded Commerce’s decision disallowing the amortization of R&D expenses and its unwillingness to accept Hy-nix’s product-specific R&D expenses.
On remand for the second time, unable to offer further support, Commerce recalculated the duty to 2.07%, although it maintained its objections for the purposes of appeal. Final Results of Redetermination Pursuant to Court Remand: Hynix Semiconductor, Inc., Hynix Semiconductor America, Inc. v. the United States and Micron Technology, Inc., (Court No. 01-00988) (Dec. 17, 2003). This second remand decision was affirmed in Hynix III.
Hynix appealed the court’s judgments to this court, claiming that: (1) Commerce’s decision to use a sales-based methodology to determine the dumping duty was inaccurate; (2) Commerce’s decision disallowing the indefinite deferral of certain R&D costs was error; and (3) Commerce’s decision disallowing the use of fixed asset revaluation to offset foreign currency translation losses was error. Commerce and Micron cross-appealed the court’s judgment validating Hynix’s amortization of R&D costs and use of product-specific R&D costs. We exercise jurisdiction pursuant to 28 U.S.C. § 1295(a)(5).
Discussion
“We review the Court of International Trade’s judgment, affirming or reversing the final results of an administrative review, de novo.” Fag Kugelfischer Georg Schafer AG v. United States, 332 F.3d 1370, 1372 (Fed.Cir.2003). In so doing, “[w]e apply anew the same standard used by the court, and will uphold Com*1368merce’s determination unless it is unsupported by substantial evidence on the record, or otherwise not in accordance with law.” Yancheng Baolong Biochemical Prods. Co. v. United States, 337 F.3d 1332, 1333 (Fed.Cir.2003) (citation and quotation marks omitted).
I. Sales-Based Methodology
We turn first to Hynix’s contention that Commerce should have calculated the duty based on a hybrid of sales and entries. Hynix has an uphill battle because “Commerce is the ‘master of anti-dumping law,’ and reviewing courts must accord deference to the agency in its selection and development of proper methodologies.” Thai Pineapple Pub. Co. v. United States, 187 F.3d 1362, 1365 (Fed.Cir.1999).
Hynix argues that a hybrid methodology is preferred, even required, by 19 U.S.C. § 1675(a)(2)5 because it is more accurate; and that, while Commerce has often used a sales-based methodology, it deviates when circumstances warrant. According to Hy-nix, the circumstances of this case warrant such a departure because consistency, Commerce’s principal desire, is irrelevant in a final review. Further, Hynix claims that it has provided sufficient data to use a hybrid sales-plus-entries methodology as illustrated by Commerce’s use of such a methodology in the Preliminary Results.
Commerce counters that section 1675(a)(2)6 requires only that the duty apply to all entries; the statute does not, however, specify how the duty should be calculated. As a result, Commerce argues that it has substantial latitude in choosing a methodology. Commerce also notes that 19 C.F.R. § 351.213(e)7 specifically empowers it to use a sales-based methodology. Further, it is Commerce’s standard practice to use a sales-based methodology, and it has in fact used such a methodology in the previous reviews of this merchandise. And, while Hynix claims to have sufficient data to support the use of entries, it was unable to match the sales of all entries because some entries were held in inventory until long after the POR ended. Finally, Commerce explained its use of a hybrid methodology in the Preliminary Results and its reversion to a sales-based methodology in the Final Results. Basically, Commerce had no justification for departing in the Preliminary Results from its established practice of using a sales-based methodology.
Commerce has offered substantial evidence supporting its choice of methodology. Section 1675(a)(2) does not mandate a specific methodology for computing the dumping margin, and section 351.213(e) specifically allows the use of “entries, exports, or sales.” See Allegheny Ludlum Corp. v. United States, 346 F.3d 1368, 1373 (Fed.Cir.2003) (holding that section 1675(a)(2) does not indicate how Commerce should calculate the dumping margin); Ad Hoc Comm. of S. Cal. Producers of Gray Portland Cement, 914 F.Supp. 535, 544 (Ct. Int’l Trade 1995) (listing Commerce’s reasons for using a sales-based methodology). Even assuming that *1369Hynix provided the necessary data linking entries and sales, Commerce need not adopt a hybrid methodology in place of its usual sales-based methodology. See Allegheny, 346 F.3d at 1371-72 (holding that “ ‘nothing in Commerce’s regulations supports the use of a hybrid sales plus [entries] approach for calculating dumping margins’ ” (quoting Hynix I at 1304)); cf. Thai Pineapple, 187 F.3d at 1365 (“Even if it is possible to draw two inconsistent conclusions from evidence contained in the record, this does not mean that Commerce’s findings are not supported by substantial evidence.”). “The methodologies relied upon by Commerce in making its determination are presumptively correct,” id., and Hynix has failed either to undermine Commerce’s rationale or to support its own position. We therefore affirm the use of a sales-based methodology in this review.
II. Research & Development
We turn next to Commerce’s treatment of Hynix’s R&D expenses, which requires a clarification of 19 U.S.C. § 1677b(f)(l)(A), which, in pertinent part, states that, “Costs shall normally be calculated based on the records of the exporter or producer of the merchandise, if such records are kept in accordance with the generally accepted accounting principles of the exporting country (or the producing country, where appropriate) and reasonably reflect the costs associated with the production and sale of the merchandise.” We have interpreted this statute in several previous cases. For example, in Thai Pineapple, 187 F.3d at 1366, we held that “[a]s a general rule, an agency may either accept financial records kept according to generally accepted accounting principles in the country of exportation, or reject the records if accepting them would distort the company’s true costs.” See also Am. Silicon Techs. v. United States, 261 F.3d 1371, 1377 (Fed.Cir.2001); NTN Bearing Corp. v. United States, 74 F.3d 1204, 1206 (Fed.Cir.1995) (“[The government] may accept [a company’s] records, but it also may reject those records if accepting them would distort the company’s true costs.”). The statute and our prior pronouncements are clear. The company has the responsibility of showing that its records are kept in accordance with its home country’s GAAP. If the company meets this burden, Commerce may counter with substantial evidence that the records do not comply with the home country’s GAAP. If the records withstand this scrutiny, Commerce may show, by substantial evidence, that the costs do not reasonably reflect the costs of production and should not, therefore, be used. See Thai Pineapple, 187 F.3d at 1367 (“If the records are not reasonably reflective of cost, Commerce may appropriately deviate from them.”). This case presents three scenarios which help to clarify the application of section 1677b(f) further.
A. Amortization vs. Expensing
During the less-than-fair value investigation, Hynix amortized its R&D expenses.8 See Dynamic Random Access Memory Semiconductors of One Megabit or Above from the Republic of Korea, 58 Fed. Reg. 27,520 (May 10, 1993). Hynix switched methods, however, in the first review, and began expensing its R&D costs.9 In the year that Hynix changed from amortizing to expensing, its R&D *1370costs were unusually high. As a result, the Court of International Trade cautioned Commerce not to include all expenses reported in that year in the dumping calculation because it would not fairly represent the cost of production. Micron Tech., Inc. v. United States, 44 F.Supp.2d 216, 221 (Ct. Int’l Trade 1999). In the fifth review, Hynix reverted to amortizing. To avoid the effect of changing methods, Commerce expensed Hynix’s R&D for all reviews, including the fifth, sixth, and final reviews. Hynix did not challenge this practice until the final review, however. Now it claims that its choice to amortize (as well as its choice to switch methods) complies with Korean GAAP and that Commerce verified its records; therefore Commerce is required by section 1677b(f) to accept its amortized R&D expenses.
Commerce acknowledges that Hynix’s decision to change accounting methods complies with Korean GAAP. Nevertheless, Commerce argues that it need not accept a company’s records pursuant to section 1677b(f) if they do not represent the costs of producing the merchandise. In this case, Commerce argues that the change between methods has resulted in the underreporting of R&D expenses. As Commerce explained, underreporting occurs when a company switches from ex-pensing to amortizing. Instead of recognizing one-fifth of the current year’s R&D costs in addition to one-fifth from the previous four years, Hynix only recognized one-fifth during this POR because the pri- or years’ costs had been expensed. Commerce also noted that Hynix was unable to show that it had used the questioned accounting methods historically. See Am. Silicon, 261 F.3d at 1379 (“19 U.S.C. § 1677b(f)(l)(A) states that Commerce shall consider the exporter’s or producer’s records on allocation of costs only if the records of the exporter or producer have been historically used by the exporter or producer.”); Statement of Administrative Action, Uruguay Round Agreements Act, Pub. L. No. 103-465, 103rd Cong.2d Sess., H. Doc. 103-316, vol. I, at 834 (1994) (“The exporter or producer will be expected to demonstrate that it has historically utilized such allocations, particularly with regard to the establishment of appropriate amortization and depreciation periods and allowances for capital expenditures and other development costs.”).
While Hynix was able to show that its records comply with Korean GAAP, Commerce showed by substantial evidence that Hynix’s reported R&D expenses fail to reflect the costs of production. It is facially apparent that a fraction of costs does not accurately capture full costs. Even if the inadequacy of this method were not transparent, it would be appropriate for us to defer to Commerce’s judgment that Hy-nix’s change from amortizing to expensing and back again results in underreporting. See Thai Pineapple, 187 F.3d at 1367 (“Antidumping investigations are complex and complicated matters in which Commerce has particular expertise and thus, Commerce’s determinations are entitled to deference.”). Further, Hynix was not able to demonstrate that it historically used this accounting methodology, which undermines the assumption that its bookkeeping was not influenced by a desire to minimize costs during the review. The court’s judgment requiring Commerce to accept Hy-nix’s amortized R&D expenses is reversed. The case is remanded to the court with instructions to remand to Commerce, which shall recalculate the duty by expens-ing Hynix’s R&D costs as in the Final Results.
B. Product-Specific R&D Expenses
We turn next to that part of the court’s judgment requiring Commerce to use Hynix’s verified, product-specific R&D costs. In the Final Results, Commerce rejected Hynix’s product-specific R&D costs in favor of a method that allocated all *1371semiconductor R&D expenses over the total number of semiconductors sold. Commerce claimed that this approach more accurately captures R&D expenses attributable to the subject merchandise because, according to the theory of cross-fertilization, other semiconductor R&D benefits DRAMs. Commerce was able to offer only two pieces of support for this assertion. First, it relied on the expertise of Dr. Murzy Jhabvala, a semiconductor engineer, who conducted a study showing that R&D for one type of semiconductor influenced the development of SRAMs. Dr. Jhabvala extrapolated from this study, without assessing Hynix’s operations specifically, that all of Hynix’s semiconductor research benefits DRAMs. Second, aside from this theoretical assessment of Hynix’s R&D environment, Commerce relied on Exhibit 20 of the Department’s verification report, which listed R&D projects in Hy-nix’s non-memory labs.10 The list contained names that appeared to indicate that Hynix’s non-memory labs were being used for memory research, even research on DRAMs specifically. Hynix, on the other hand, argues that Commerce must accept its verified, product-specific R&D costs pursuant to section 1677b(f), and, further, that section 1677b(e)(2)(A)11 requires Commerce to use actual values, such as those provided by Hynix, when available.
This precise issue was presented to the court in a prior review of the same merchandise. In Micron Technology, Inc. v. United States, 893 F.Supp. 21, 27 (Ct. Int’l Trade 1995), aff'd, 117 F.3d 1386 (Fed.Cir.1997), the court required Commerce to accept the product-specific R&D expenses reported by Hynix. In so doing, it noted that “the validity of Commerce’s methodology cannot rest on intuitive appeal alone; rather, the factual premise upon which Commerce bases its choice of methodology must be supported by substantial evidence on the record.” 893 F.Supp. at 27. We agree with this assessment and adopt it as our own. Dr. Jhabvala’s opinion, while probative of the company he investigated or potentially SRAMs generally, bears little relation to either Hynix or DRAMs. And, as noted by Hynix, the list of projects cited by Commerce is paltry evidence of cross-fertilization, because project names, without proof of the underlying activities, do little to show that all semiconductor R&D has impacted the development of DRAMs.
Commerce has failed to offer “such relevant evidence as a reasonable mind might accept as adequate to support” the proposition that Hynix’s product-specific R&D costs are not representative. NTN Bearing, 74 F.3d at 1206 (internal quotation marks and citation omitted). The court’s judgment requiring the use of Hynix’s product-specific R&D expenses is affirmed.
C. Indefinite Deferral of R&D Expenses
In the Final Results, Commerce refused to allow the indefinite deferral of certain R&D expenses. Hynix argues that this violated section 1677b(f) because its records were verified and deferral is allowed under Korean GAAP and International Accounting Standard 9 (“IAS”).12 *1372According to Hynix, the matching principle13 requires the indefinite deferral of these expenses because associated revenues will not be realized until some time in the future. As evidence, it cites to a single document that notes, almost illegibly, that certain projects will mature sometime in 2001 or 2002.
Commerce argues that the matching principle does not apply because “there is no objective evidence or reasonable expectation that deferred R&D costs will result in future benefits” as required by both Korean GAAP and IAS 9. Remand, I at 9. According to Commerce, “[ajccountants defer recognition of an expense to the future only when there is reasonable evidence that the expenditure will, in fact, benefit future operations. If this evidence is not available, or is not convincing, accountants do not attempt to apply the matching principle; rather they charge the expenditure immediately to expense.” Id. at 8 (quoting Robert F. Meigs & Walter B. Meigs, Financial Accounting 734 (7th ed.1992)). Pursuant to this approach, Commerce claims that the documentation offered by Hynix is insufficient and, therefore, conservatism mandates expensing these costs.
While both Korean GAAP and IAS 9 allow the deferral of R&D expenses, neither standard allows deferral unless certain conditions are satisfied, namely that there be a reasonable expectation of future benefit. Commerce concluded, and we accept, that Hynix offered insufficient evidence that its R&D costs will result in future revenues. See Am. Silicon, 261 F.3d at 1380 (“This court ‘accords deference to the determinations of the agency that turn on complex economic and accounting inquiries.’ We hold that determining whether a depreciation practice comports with Brazilian GAAP is one such complex economic and accounting inquiry.” (citation omitted)). Because Hynix’s records do not comport with the requirements of its home country’s GAAP, Commerce need not use them to calculate the duty. The court’s decision affirming the disallowance of the indefinite deferral of certain R&D costs is affirmed.
III. Translation Losses
Finally, we address Hynix’s contention that Commerce should have offset its foreign currency translation losses14 with the revaluation of its fixed assets. In 1997, the Korean won steeply declined in value, necessitating the revaluation of Hynix’s fixed assets. It is undisputed that this revaluation was conducted in accordance with the Asset Revaluation Law of Korea and Korean GAAP.
According to Hynix, its translation losses were incurred on the very loans used to purchase the fixed assets that were revalued, and that the same phenomenon, the devaluation of the won, caused both. Hy-nix contends, therefore, that it is consistent with the principle of matching to offset the translation losses with the increase in value of its fixed assets. Hynix also complains that Commerce’s method actual*1373ly double-counts the effects of the devaluation of the won because depreciation costs grew, as the assets increased in value, and translation costs increased.
Neither Commerce nor the court accepted Hynix’s arguments. In fact, Hynix has attempted to make these same arguments in the past and failed. See Micron Tech, 893 F.Supp. at 33. As noted by Commerce in the unpublished decision memorandum accompanying its Final Residís, “revaluation of the fixed assets (i.e., non-monetary assets) does not represent income during the year. In economic terms, the company is in the same position holding the same assets.” We accept this characterization both because it is correct and because we owe Commerce deference when it makes this type of technical decision. See Am. Silicon, 261 F.3d at 1380-81. It is obvious, therefore, that Hynix’s arguments are without merit. That the revaluation does not represent income is enough to show that offsetting is inappropriate — one cannot offset a real loss15 with an imagined gain. Hynix itself did not treat the translation losses and revaluation in its own books as it would have Commerce treat them. See 19 U.S.C. § 1677b(f) (2000). The refusal to offset Hynix’s foreign currency translation losses with the revaluation of its fixed assets is affirmed.
Conclusion
Accordingly, the judgments of the Court of International Trade are affirmed-in-part, reversed-in-part and the case is remanded for further proceedings in accordance with this opinion.
COSTS
No costs.
AFFIRMED IN-PART, REVERSED IN-PART AND REMANDED
. The underlying administrative review was initiated against Hyundai Industries Co., Ltd. and LG Semicon Co., Ltd. Hyundai purchased LG Semicon during the Period of Review and changed its name to Hynix Semiconductor, Inc.
. The seventh review, which was originally slated to expire on April 30, 2000, was truncated by Commerce’s decision that it was unlikely that future dumping would occur if the dumping order was revoked. See Dynamic Random Access Memory Semiconductors of One Megabit or Above From the Republic of Korea; Final Results of Full Sunset Review and Revocation of Order, 65 Fed. Reg. 59,391 (Oct. 5, 2000).
. Conservatism teaches that costs should be expensed or realized in the year incurred.
. According to Commerce, "[c]ross-fertilization is [a theory] based upon the widely recognized notion in the semiconductor industry, that research in one area benefits research in other areas, i.e. memory research benefits non-memory research and vice versa." Remand I at 10.
. Hynix specifically relies on subsection (A), which states that Commerce "shall determine — (i) the normal value and export price (or constructed export price) of each entry of the subject merchandise, and (ii) the dumping margin for each such entry.”
. Commerce specifically relies on subsection (C), which states that "[t]he determination under this paragraph shall be the basis for the assessment of countervailing or antidumping duties on entries of merchandise covered by the determination and for deposits of estimated duties.”
.19 C.F.R. § 351.213(e)(l)(i) states that "an administrative review under this section normally will cover, as appropriate, entries, exports, or sales of the subject merchandise.” (emphasis added).
. By way of example, when a company amortizes costs, it recognizes a fraction of its costs, one-fifth in this case, in the year incurred. In the following four years the remaining four-fifths are recognized, one-fifth per year.
. When, a company expenses, it recognizes all R&D costs in the year incurred. All costs not recognized as the result of prior amortization are recognized in that year as well.
. The following R&D projects were listed: "[Embedded SRAMs],” "[Embedded DRAMs],” "[DDR Memory Resister],” "[Smart Memory]” "[MPEG2 + DRAM],” "[Memory Register],” and "[DRAM Macro].” Remand I at 11.
. Section 1677b(e)(2)(A) states that "the constructed value of imported merchandise shall be an amount equal to the sum of — ... (2)(A) the actual amounts incurred and realized by the specific exporter or producer being examined in the investigation or review .... ” (emphasis added).
.IAS 9 states that "development costs are recognized as an asset when they meet certain criteria that indicate that it is probable *1372that the costs will give rise to future economic benefits.”
. The matching principle teaches that costs should be realized in the same period as associated benefits. Under this theory, an expenditure in 2000 that does not generate revenues until 2001 should be expensed in 2001— not in 2000 when the cost was actually incurred.
. Foreign currency translation losses are the costs associated with holding dollar-denominated debt. When the home country's currency drops in value, dollar-denominated debt becomes more costly to pay (i.e., it takes more won to pay off dollar-denominated debt when the value of the won decreases).
. There is ample evidence that translation losses are real. “Significantly, Korean GAAP reflects as much by requiring that translation losses be recognized in the year in which they are incurred.” Micron Tech., 893 F.Supp. at 33.